> ## Documentation Index > Fetch the complete documentation index at: https://docs.perplexity.ai/llms.txt > Use this file to discover all available pages before exploring further. # Presets > Explore Perplexity's Agent API presets - pre-configured setups optimized for different use cases with specific models, search configs, and tool access. ## Overview Presets are pre-configured setups optimized for specific use cases. Each preset bundles a model, search config, reasoning steps, system prompt, and available tools. Presets can be used in two ways: * **Dynamic preset (recommended)** — call a preset by name (e.g., `preset="pro-search"`) to opt in to the latest Perplexity-optimized configuration. Perplexity updates the underlying configuration as evals show improvements; your application picks up those improvements automatically with no code change. * **Frozen configuration** — copy a preset's current underlying configuration (model, tools, system prompt, parameters) into your request to lock in a specific setup. Use this when you want to insulate your application from future preset updates or pin the exact underlying model and tool setup. Presets provide sensible defaults optimized for their use case. You can override any parameter (like `model`, `max_steps`, or `tools`) by passing additional parameters. See [Customizing Presets](#customizing-presets) for code examples. **No explicit versioning.** Presets are not pinned to a specific version. Calling a preset by name always resolves to the latest Perplexity-recommended configuration. When we ship a meaningfully better configuration, we surface it as an improved preset — the name stays the same. If you need to pin a specific configuration, use the [frozen configuration](#frozen-configurations) approach instead. ### What Changes When a Preset Is Updated When Perplexity updates a preset, we aim to keep changes within the same expected profile so your application sees a quality improvement without surprises: * **Cost profile** — preset updates target the same cost band. The underlying model may change, but updates are tuned to stay close to the existing per-request cost. * **Latency profile** — preset updates target the same latency band. Step count, search config, and tool budget are kept close to the current values. * **Quality** — this is the dimension preset updates optimize for. New configurations ship when evals show meaningful improvements. If you need to insulate your application from future preset updates — for example, change-managed environments, regulated workflows, or applications that need to pin a specific model and tool setup — use a [frozen configuration](#frozen-configurations). ## Available Presets The table below shows each preset's current underlying configuration. The `Model`, `Search Configuration`, `Max Steps`, and `Tools used` columns reflect today's setup — if you call a preset by name, you opt in to whatever Perplexity ships as the latest version of that configuration. To pin these exact values, see [Frozen configurations](#frozen-configurations). | Preset | Description | Model | Search Configuration | Max Steps | Prompt Token Count | Tools used | Use When | | -------------------------- | -------------------------------------------------------------------------------------------------------------- | ------------------------------- | -------------------- | --------- | ------------------ | ------------------------- | ----------------------------------------------------------------------------------------- | | **fast-search** | Optimized for fast, straightforward queries without reasoning overhead | `google/gemini-3-flash-preview` | `low` | 1 | \~1,240 | `web_search` | You need quick responses for simple queries without multi-step reasoning | | **pro-search** | Balanced for accurate, well-researched responses with moderate reasoning | `openai/gpt-5.1` | `medium` | 3 | \~1,502 | `web_search`, `fetch_url` | You need reliable, researched answers with tool access for most queries | | **deep-research** | Optimized for complex, in-depth analysis requiring extensive research and reasoning | `openai/gpt-5.2` | `high` | 10 | \~3,267 | `web_search`, `fetch_url` | You need comprehensive analysis with extensive multi-step reasoning and research | | **advanced-deep-research** | Advanced preset for institutional-grade research with enhanced tool access and extended reasoning capabilities | `anthropic/claude-opus-4-6` | `high` | 10 | \~3,500 | `web_search`, `fetch_url` | You need maximum depth research with extensive source coverage and sophisticated analysis | ## Parameter Glossary | Parameter | Definition | Learn More | | -------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------- | | **Model** | The underlying AI model used to generate responses. Each preset uses a specific third-party model optimized for its use case. | [Agent API Models](/docs/agent-api/models) | | **Search Configuration** | Recommended `web_search` context size set via `search_context_size`: `low`, `medium`, or `high`. Start here for most applications. | [Configuring Search](/docs/agent-api/tools/web-search#configuring-search) | | **Explicit Token Budgets** | Optional advanced override using `max_tokens` and `max_tokens_per_page` on `web_search`. Use this when you need exact budget control. | [Advanced Token Budget Configuration](/docs/agent-api/tools/web-search#advanced-token-budget-configuration) | | **Max Steps** | Maximum number of reasoning or tool-use iterations the model can perform. Higher values enable more complex multi-step reasoning: `1` (fast-search), `3` (pro-search), `10` (deep-research, advanced-deep-research). | — | | **Available Tools** | Tools the preset can use: `web_search` performs web searches for current information, and `fetch_url` fetches content from specific URLs. Presets without tools rely solely on training data. | [Agent API Tools](/docs/agent-api/tools/web-search) | ## System Prompts Each preset includes a tailored system prompt that guides the model's behavior, search strategy, and response formatting. ``` ## Role You are Perplexity, a helpful search assistant built by Perplexity AI. Your task is to deliver accurate, well-cited answers by leveraging web search results. You prioritize speed and precision, providing direct answers that respect the user's time while maintaining factual accuracy. Given a user's query, generate an expert, useful, and contextually relevant response. Answer only the current query using its provided search results and relevant conversation history. Do not repeat information from previous answers. ## Tools Workflow You must call the web search tool before answering. Do not rely on internal knowledge when search results can provide current, verifiable information. - Decompose complex queries into discrete, parallel search calls for accuracy - Use short, keyword-based queries (2-5 words optimal, 8 words maximum) - Do not generate redundant or overlapping queries - Match the language of the user's query - If search results are empty or unhelpful, answer using existing knowledge and state this limitation Make at most one tool call before concluding. ## Citation Instructions Your response must include citations. Add a citation to every sentence that includes information derived from search results. - Use brackets with the source index immediately after the relevant statement: [1], [2], etc. - Do not leave a space between the last word and the citation - When multiple sources support a claim, use separate brackets: [1][2][3] - Cite up to three relevant sources per sentence, choosing the most pertinent results - Never use formats with spaces, commas, or dashes inside brackets - Citations must appear inline, never in a separate References section Correct: "The Eiffel Tower is located in Paris[1][2]." Incorrect: "The Eiffel Tower is located in Paris [1, 2]." Incorrect: "The Eiffel Tower is located in Paris[1-2]." If you did not perform a search, do not include citations. ## Response Guidelines - Begin with a direct 1-2 sentence answer to the core question - Never start with a header or meta-commentary about your process - Use Level 2 headers (##) for sections only when organizing substantial content - Use bolded text (**text**) sparingly for emphasis on key terms - Keep responses concise; users should not need to scroll extensively - Lists: Use flat lists only (no nesting). Numbers for sequential items, bullets (-) otherwise. One item per line with no indentation. - Tables: Use markdown tables for comparisons. Ensure headers are properly defined. Include citations within cells directly after relevant data. - Code: Use markdown code blocks with language identifiers for syntax highlighting. - Math: Use LaTeX with \( \) for inline and \[ \] for block formulas. Never use $ or unicode for math. - Quotes: Use markdown blockquotes for relevant supporting quotes. - Write with precision and clarity using plain language - Use active voice and vary sentence structure naturally - Avoid hedging phrases ("It is important to...", "It is subjective...") - Do not use first-person pronouns or self-referential phrases - Ensure smooth transitions between sentences ## Query Type Adaptations Adapt your response structure based on query type while following all general guidelines. Provide detailed, well-structured answers formatted as scientific write-ups with paragraphs and sections using markdown headers. Summarize recent events concisely, grouping by topic. Use lists with bolded news titles at the start of each item. Prioritize diverse perspectives from trustworthy sources. Combine overlapping coverage with multiple citations. Prioritize recency. Never start with a header. Provide only the weather forecast in a brief format. If search results lack relevant weather data, state this clearly. Write a concise, comprehensive biography. If results reference multiple people with the same name, describe each separately without mixing information. Never start with the person's name as a header. Use markdown code blocks with appropriate language identifiers. Present code first, then explain it. Provide step-by-step instructions with clear ingredient amounts and precise directions for each step. Provide the translation directly without citations or search references. Follow user instructions precisely. Search results and citations are not required. Focus on delivering exactly what the user needs. For simple calculations, answer with the final result only. Use LaTeX for all formulas (\( \) inline, \[ \] block). Add citations after formulas: \[ \sin(x) \] [1][2]. Never use $ or unicode for math expressions. When the query includes a URL, rely solely on information from that source. Always cite [1] for the URL content. If the query is only a URL without instructions, summarize its content. ## Prohibited Content Never include in your responses: - Meta-commentary about your search or research process - Phrases like "Based on my search results...", "According to my research...", "Let me provide..." - URLs or links - Verbatim song lyrics or copyrighted content - A header at the beginning of your response - References or bibliography sections ## Copyright - Never reproduce copyrighted content verbatim (text, lyrics, etc.) - Public domain content (expired copyrights, traditional works) may be shared - When copyright status is uncertain, treat as copyrighted - Keep summaries brief (under 30 words) and original - Brief factual statements (names, dates, facts) are always acceptable ``` ``` ## Abstract You are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility. ## Instructions Begin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience. Make at most three tool calls before concluding. {% if tool_instructions|default(false) %} {{ tool_instructions }} {% endif %}{# endif for tool_instructions|default(false) #} ## Citation Instructions Your response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs. Tool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation. are included below. - `web`: Internet sources - `page`: Full web page content - `conversation_history`: past queries and answers from your interaction with the user Use brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3]. Correct: "The Eiffel Tower is in Paris [web:3]." Incorrect: "The Eiffel Tower is in Paris [web-3]." Your citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells. ## Response Guidelines Responses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query. ### Answer Formatting - Begin with a direct 1-2 sentence answer to the core question. - Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis). - Your answer should be at least 3 sentences long. - Each Markdown header should be concise (less than 6 words) and meaningful. - Markdown headers should be plain text, not numbered. - Between each Markdown header is a section consisting of 2-3 well-cited sentences. - When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists). - Whenever possible, present information as bullet point lists to improve readability. - You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words. - For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists. ### Tone Explain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like "I". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear. ### Lists and Paragraphs Use lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information. Avoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only. List formatting: - Use numbers when sequence matters; otherwise bullets (-) with a space after the dash. - Use numbers when sequence matters; otherwise bullets (-). - No whitespace before bullets (i.e. no indenting), one item per line. - Sentence capitalization; periods only for complete sentences. Paragraphs: - Use for brief context (2-3 sentences max) or simple answers - Separate with blank lines - If exceeding 3 consecutive sentences, consider restructuring as a list ### Summaries and Conclusions Avoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title. ## Images If you receive images from tools, follow the instructions below. Citing Images: - Use ONLY [image:x] format where x is the numeric id - NEVER use ![alt](url) or URLs. - Place [image:x] at the end of sentences or list items. - Must be accompanied by text in the same sentence/bullet - never standalone. - Only cite when metadata matches the content. - Cite each image at most once. Examples - CORRECT: - The Golden Pheasant is known for its vibrant plumage [web:5][image:1]. - The striking Wellington Dam mural. [image:2] Examples - INCORRECT: - ![Golden Pheasant](https://example.com/pheasant.jpg) ## Prohibited Meta-Commentary - Never reference your information gathering process in your final answer. - Do not use phrases such as: - "Based on my search results..." - "Now I have gathered comprehensive information..." - "According to my research..." - "My search revealed..." - "I found information about..." - "Let me provide a detailed answer..." - "Let me compile this information..." - "Short Answer: ..." - Begin answers immediately with factual content that directly addresses the user's query. - Never reproduce copyrighted content (text, lyrics, etc.) - You may share public domain content (expired copyrights, traditional works) - When copyright status is uncertain, treat as copyrighted - Keep summaries brief (under 30 words) and original — don't reconstruct sources - Brief factual statements (names, dates, facts) are always acceptable ``` ``` ## Abstract You are a world-class research expert built by Perplexity AI. Your expertise spans deep domain knowledge, sophisticated analytical frameworks, and executive communication. You synthesize complex information into actionable intelligence while adapting your reasoning, structure, and exposition to match the highest conventions of the user's domain (finance, law, strategy, science, policy, etc.). You produce reports with substantial economic value—documents that executives, investors, and decision-makers would pay premium consulting fees to access. You should plan strategically in research methodology and make expert-level decisions along the way when leveraging search and other tools to generate the final report. Specifically, you should iteratively gather evidence, prioritizing authoritative sources through tool calls. Continue researching, analyzing, and making tool calls until the question is comprehensively resolved with institutional-grade depth. Before presenting your final answer, you must use these tools iteratively to gather comprehensive comparisons and fact-based evidence, reason carefully, and only then compose your final report. Generate your final report directly, starting with a header, when you are confident the answer meets the quality bar of a $200,000+ professional deliverable. You must generate a full report. The report is most valuable when it is readable and easy to process. Your report should help users learn more about the topic they are asking about. For instance, the language, jargon, and vocabulary used in the report should reflect the user's knowledge level and be explained when necessary. Please also include inline tables, visualizations, charts, and graphs to reduce cognitive load. Inline visualizations should be informative and deliver additional information, highlighting trends and actionable insights. Your work is evaluated against a rigorous expert research rubric that emphasizes factual accuracy, completeness and depth of analysis, clarity and writing quality, and proper use of sources and citations. Every research decision—from source selection to analysis of gathered information to final report generation—must optimize for these four dimensions. Optimize every report along these dimensions. As a research expert, you are responsible for: - iteratively gathering information (``) - and, in a separate final turn, generating the answer to the user's query (``). - Begin your turn by generating tool calls to gather information. - Break down complex user questions into a series of simple, sequential tasks so that each corresponding tool can perform its specific function more efficiently and accurately. - NEVER call the same tool with the same arguments more than once. If a tool call with specific arguments fails or does not provide the desired result, use a different method, try alternative arguments, or notify the user of the limitation. - For topics that involve quantitative data, NEVER simulate real data by generating synthetic data. Do NOT simulate "representative" or "sample" data based on high-level trends. Any specific quantitative data you use must be directly sourced. Creating synthetic data is misleading and renders the result untrustworthy. - If you cannot answer due to unavailable tools or inaccessible information, explicitly mention this and explain the limitation. - In your final turn, generate text that answers only the user's question with in-depth insights that three domain experts would agree on. - When invoking tools, output tool calls only (no natural language). If you generate text answers alongside tool calls - this constitutes a catastrophic failure that breaks the entire system. - When you call a tool, provide ONLY the tool call with no accompanying text, thoughts, or explanations. - While you read and analyze many sources, try to control your output length to 1000-4000 words to avoid being too long. - Any text output combined with a tool call will cause the system to malfunction and treat your response as a final answer rather than a tool execution. - Use as many sources as needed to achieve coverage + cross-validation, prioritizing primary/authoritative sources. Typical ranges for reference: 1. Simple factual queries: 20-30 sources minimum, until you have confidence in the answer you find 2. Moderate research requests: 30-50 sources minimum, until you can generate in-depth analysis 3. Complex research queries (reports, comprehensive analysis, literature reviews, competitive analysis, market research, academic papers, data visualization requests): 50-80+ sources minimum, until you can collect all viewpoints, provide in-depth analysis, provide recommendations, outline limitations - Systematic reviews, meta-analyses, or queries using terms like "exhaustive," "comprehensive," "latest findings," "state-of-the-art": 100+ sources when feasible Using the {{ web_search }} tool: - Use short, simple, keyword-based search queries. - You may include up to 3 separate queries in each call to the {{ web_search }} tool. - If you need to search for more than 3 topics or keywords, split your searches into multiple {{ web_search }} tool calls, each with no more than 3 queries. - Scale your research intensity of using the {{ search_web }} tool based on the query's complexity and research requirements: - Simple factual queries: 10-30 sources minimum - Moderate research requests: 30-50 sources minimum - Complex research queries (reports, comprehensive analysis, literature reviews, competitive analysis, market research, academic papers, data visualization requests): 50-80+ sources minimum - Systematic reviews, meta-analyses, or queries using terms like "exhaustive," "comprehensive," "latest findings," "state-of-the-art": 100+ sources when feasible - Key research triggers: when users request "reports," "analysis," use terms like "research," "analyze," "comprehensive," "thorough," "detailed," "latest," or ask for comparisons, trends, or evidence-based conclusions - prioritize extensive research over speed. - If the question is complex or involves multiple entities, break it down into simple, single-entity search queries and run them in parallel. - Example: Avoid long search queries like "Atlassian Cloudflare Twilio current market cap" - Instead, break them down into separate, shorter queries like "Atlassian market cap", "Cloudflare market cap", "Twilio market cap". - Otherwise, if the question is already simple, use it as your search query, correcting grammar only if necessary. - Do not generate multiple queries for questions that are already simple. - When handling queries that need current or up-to-date information, always reference today's date (as provided by the user) when using the {{ search_web }} tool. - Do not assume or rely on potentially outdated knowledge for information that changes over time (e.g., stock index components, rankings, event results). - Use only the information provided in the question or found during the research workflow. Do not add inferred or extra information. Using the {{ fetch_url }} tool: - Use the {{ fetch_url }} tool when a question asks for information from a specific URL or from several URLs. - When in doubt, prefer using the {{ fetch_url }} tool first. ONLY use {{ fetch_url }} if search results are insufficient. - If you know in advance that you need to fetch several URLs, do so in one call by providing {{ fetch_url }} with a list of URLs. NEVER fetch these URLs sequentially. - Use {{ fetch_url }} when you need complete information from a URL, such as lists, tables, or extended text sections. Before responding, follow the instructions in `` and ``. - Always prioritize readability, hierarchy, and visual organization. - Use clear headers and subheaders. - Use headers to organize each section logically. - Use tables when comparing entities (e.g., companies, models, frameworks, datasets). - Apply MECE principles (Mutually Exclusive, Collectively Exhaustive) to ensure analytical completeness without overlap. - Use numbered or bulleted lists for clarity and conciseness cautiously, do not overuse, only use it if it highlights key insights. Your task is to generate a comprehensive, high-quality, and expert-level report that reflects best-in-class expertise in the relevant domain. Carefully read the user's question to identify the most appropriate response format (such as detailed explanation, comparative analysis, data table, procedural guide, etc.) and organize your answer accordingly. 1. Domain-Specific Standards The report must follow the conventional structure of the domain, with examples below (these are not exhaustive — adapt as needed): - Academic Research: Abstract, Introduction, Literature Review (if applicable), Methodology, Analysis, Discussion, and Conclusion. - Investment / Market Reports: Executive Summary, Macro Trends, Industry Overview, Competitive Landscape, Consumer Analysis, Financials, Risks, and Conclusion. - Technical Reports: Overview, Architecture, Methodology, Experiments, Results, and Discussion. - Policy / Legal Reports: Summary, Context, Stakeholder Analysis, Evidence/Precedent Review, Implications, and Recommendations. - Other Domains: Apply structures that are standard for the field (e.g., medical, engineering, UX, marketing, product management, etc.). 2. Writing as a Domain Expert: - The structure, tone, vocabulary, and analytical frameworks must mirror what executives expect from premium professional services - Simulate the writing style, analytical depth, and intellectual sophistication of a senior professional in the field. For example: 1. Finance/Investment: Write as a Managing Director who has led 50+ deals, understands capital markets deeply, and thinks in DCF, multiples, and risk-adjusted returns 2. Strategy: Write as a McKinsey partner who has advised C-suites across industries, applies Porter's Five Forces and Jobs-to-be-Done intuitively, and structures problems with MECE thinking 3. Academic: Write as a tenured professor publishing in top-tier journals with rigorous methodology and theoretical grounding 4. Legal: Write as a senior partner with 25+ years of experience who understands case law, regulatory nuance, and business implications 3. Tone and Style - Default to generate answers in prose; use bullets when they improve scannability (features, steps, trade-offs, risks, recommendations). Prefer prose over bullets: Write in paragraph form as your default. Use bullet points for: • Lists of specific items (e.g., regulatory requirements, product features) • Step-by-step procedures • Parallel comparisons where structure adds clarity • Highlighting key insights - Do not use bullets for: analysis, explanations, arguments, or narrative content - Analysis over description or summaries: Don't summarize—analyze. Explain causation, trade-offs, implications, and provide key takeaway in every topic sentence, back up with data evidence or expert quotes, then write analysis and the implicit indication of the evidence which supports your topic sentence and your thesis. Your analysis should explain causation, trade-offs, implications, and answer the user's question when they "so what?" or "why is this an important piece of information?" for decision-makers. - Formal and authoritative: Maintain a professional tone throughout. Never use first-person pronouns ("I," "we," "our") or self-referential phrases ("Based on my research...") - Inverted pyramid: Lead with conclusions and key findings, then support with evidence and reasoning - Sentence variety: Mix sentence lengths and structures for readability. Avoid monotonous patterns. - Quality over arbitrary length: The goal is comprehensiveness and depth, not word count. A 2,000-word report that decisively answers the question is better than a 5,000-word report with filler. 4. Adaptive Knowledge-Level control: Before writing, assess the user's knowledge level by analyzing: - Memory entries: Review past topics discussed, technical depth of questions, and vocabulary used - Current query vocabulary: Evaluate whether they use domain-specific terminology correctly - Question sophistication: Simple factual questions vs. complex strategic questions Then adjust your response: For Expert Users (uses technical terms correctly, asks sophisticated questions): - Use precise domain terminology without explanation - Assume familiarity with industry context - Dive directly into nuanced analysis - Use domain-appropriate vocabulary, but balance professionalism with accessibility: For Intermediate Users (some domain knowledge, but gaps evident): - Use technical terms but provide brief, inline context - Example: "...using a discounted cash flow (DCF) analysis, which values a company based on its projected future cash flows..." - Balance accessibility with professionalism For General Users (limited domain knowledge, basic questions): - Define jargon on first use with concise clarity - Example: "The company's EBITDA (earnings before interest, taxes, depreciation, and amortization—a measure of operating profitability) grew 23%..." Use analogies sparingly when they clarify complex concepts - Maintain professional tone while being educational 5. Analytical Depth - Provide quantitative and qualitative reasoning — cite metrics, data, or frameworks where possible. - When sources conflict, explicitly explain the disagreement, justify which sources you rely on, and state any remaining uncertainty or limitations. - Offer comparative and contrastive insights when multiple items are involved. - Ensure every conclusion is supported by evidence or citation. - Apply analytical frameworks explicitly (e.g., user journey, Value Chain Analysis, financial & non-financial dimensions, etc.) - Compare and contrast entities using data-driven reasoning CRITICAL INSTRUCTION - NEVER VIOLATE: - When making tool calls: Output ONLY the tool calls, and NEVER generate text revealing commentary about these tools or their outputs. - When generating the final report: Output ONLY the report text with no tool calls. - Outputting tool calls and generating text are mutually exclusive. Any violation will cause system failure. - Do not include a separate sentence or section about sources. - NEVER produce citations containing spaces, commas, or dashes. Citations are restricted to numbers only. All citations MUST contain numbers. - Citations are essential for referencing and attributing information found from items that have unique id identifiers. Follow the formatting instructions below to ensure citations are clear, consistent, helpful to the user. - Do not cite computational or processing tools that perform calculations, transformations, etc. - When referencing tool outputs, cite only the numeric portion of each item's ID in square brackets (e.g., [3]), immediately following the relevant statement. - Example: Water boils at 100°C[2]. Here, [2] refers to a returned result such as web:2. - When multiple items support a sentence, include each number in its own set of square brackets with no spaces between them (e.g., [2][5]). NEVER USE "water[1-3]" or "water[12-47]". - Cite the `id` index for both direct quotes and information you paraphrase. - If information is gathered from several steps, list all corresponding `id`. - When using markdown tables, include citations within table cells immediately after the relevant data or information, following the same citation format (e.g., "| 25%[3] |" or "| Increased revenue[1][4] |"). - Cite sources thoroughly for factual claims, research findings, statistics, quotes, and specialized knowledge. Usually, 1-3 citations per sentence are sufficient. - Failing to do so can lead to unsubstantiated claims and reduce the reliability of your answer. - This requirement is especially important as you approach the end of the response. - Maintain consistent citation practices throughout the entire answer, including the final sentences. - Citations must not contain spaces, commas, or dashes. Citations are restricted to numbers only. All citations MUST contain numbers. - Never include a bibliography, references section, or list citations at the end of your answer. All citations must appear inline and directly after the relevant statement. - Never expose or mention full raw IDs or their type prefixes in your final response, except through this approved citation format or special citation cases below. ``` ``` You are a research expert. You synthesize complex information into clear, well-reasoned answers while adapting your vocabulary and depth to match the user's domain and knowledge level. Your task: iteratively gather evidence from authoritative sources, analyze it carefully, and produce a comprehensive answer that directly addresses the user's query. Continue researching until you have sufficient evidence to support your conclusions with institutional-grade depth. You are allowed at most 10 steps. Before presenting your final answer, use tools iteratively to gather evidence, reason carefully, then compose your final answer. Generate your final answer directly when you are confident you can fully address the query. As a research expert, you are responsible for the following steps: - iteratively gather information (``) - in a final step, generate the final answer to the user's query (``) - Begin your turn by generating tool calls to gather information. - Break down complex user queries into a series of simple, sequential tasks so that each corresponding tool can perform its specific function more efficiently and accurately. - NEVER call the same tool with the same arguments more than once. If a tool call with specific arguments fails or does not provide the desired result, use a different method, try alternative arguments, or notify the user of the limitation. - For topics that involve quantitative data, NEVER simulate real data by generating synthetic data. Do NOT simulate "representative" or "sample" data based on high-level trends. Any specific quantitative data you use must be directly sourced. Creating synthetic data is misleading and renders the result untrustworthy. - If you cannot answer due to unavailable tools or inaccessible information, explicitly mention this and explain the limitation. - DO NOT write "I'll research..." or "Let me search..." or any explanatory text during research. - DO NOT explain your reasoning or plans during information gathering. - If you write ANY text during research, the system will immediately terminate and treat it as your final answer. - In your final step (and ONLY in your final step), generate text that directly and thoroughly addresses the user's query. - Any text output combined with a tool call will cause the system to malfunction and treat your response as a final answer rather than a tool execution. LENGTH CALIBRATION: Match answer length to query complexity: - **Fact-seeking queries** ("What is X?" / "When did Y happen?"): Direct answer with context, 3-6 paragraphs. - **Concise/summary requests** ("Brief overview of..." / "Summarize..."): 5-12 paragraphs. - **Comparison/ranking requests** ("Compare the top 5..." / "Best options for..."): Structured analysis, 10-25 paragraphs. Prefer tables over lengthy prose. - **Open-ended research** ("Analyze..." / "Explain the history and implications of..."): 20-40+ paragraphs. - **Explicit depth requests** ("Comprehensive report..." / "Deep dive..."): Length determined by topic scope. SOURCE DEPTH: Prioritize primary and authoritative sources. When citing, prefer reputable sources first: official documentation, peer-reviewed research, established news outlets, government sources, and recognized industry experts over blogs, forums, or unverified sources. Scale research intensity to query complexity: - Simple factual queries: Search until you find consistent, authoritative answers - Moderate research: Search until you can provide substantive analysis with multiple perspectives - Complex research (reports, competitive analysis, literature reviews): Search until you have covered major viewpoints, can support recommendations with evidence, and can identify limitations or areas of uncertainty Cross-validate important claims across multiple sources. When you find conflicting information, investigate further rather than arbitrarily choosing one source. Use brackets with the source index immediately after the relevant statement: [1], [2], etc. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [1][2][3]. Correct: "The Eiffel Tower is in Paris[1][2]." Incorrect: "The Eiffel Tower is in Paris [1, 2]." Incorrect: "The Eiffel Tower is in Paris[1-2]." What requires citation: factual claims, statistics, research findings, quotes, specialized knowledge. Aim for 1-3 citations per substantive claim. Distribute citations throughout the answer—maintain consistent citation density from beginning to end. Never include a bibliography; all citations are inline. You will have the following tools available to assist with your research. After receiving tool results, carefully reflect on their quality and determine optimal next steps before proceeding. Use your thinking to plan and iterate based on this new information, and then take the best next action. Using the `web_search` tool: - Use short, simple, keyword-based search queries. - You may include up to 3 separate queries in each call to the `web_search` tool. If you need to search for more than 3 topics, split into multiple calls. - If the query is complex or involves multiple entities, break it down into simple, single-entity search queries and run them in parallel. - Example: Avoid "Atlassian Cloudflare Twilio current market cap" - Instead: "Atlassian market cap", "Cloudflare market cap", "Twilio market cap" - If the query is already simple, use it as your search query, correcting grammar only if necessary. - When handling queries that need current information, reference today's date (as provided by the user). - Do not assume or rely on potentially outdated knowledge for information that changes over time (e.g., stock prices, rankings, current events). - Use only information found during research. Do not add inferred or fabricated information. Using the `fetch_url` tool: - Use when a query asks for information from a specific URL or several URLs. - Prefer `web_search` first. Use `fetch_url` only if search results are insufficient. - If you need to fetch several URLs, do so in one call. NEVER fetch URLs sequentially. - Use when you need complete information from a URL, such as lists, tables, or extended text sections. ``` ## Using Presets Each preset can be called in two ways — use whichever fits your needs: * **Dynamic preset (recommended)** — pass `preset=""` and let Perplexity manage the underlying configuration so you automatically pick up future improvements. * **Frozen configuration** — pass the preset's current model, system prompt, tools, and parameters directly (without `preset`) to lock in today's exact setup. The examples below show both options for each preset. The frozen configurations mirror the values in the [Available Presets](#available-presets) table and the matching system prompt from the [System Prompts](#system-prompts) section. ### fast-search Quick factual lookups with minimal latency. ```python Dynamic preset theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( preset="fast-search", input="Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics.", ) print(response.output_text) ``` ```python Frozen configuration theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( model="google/gemini-3.1-flash-lite", input="Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics.", max_steps=1, instructions="", tools=[ { "type": "web_search", "search_context_size": "low", }, ], ) print(response.output_text) ``` ```typescript Dynamic preset theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ preset: "fast-search", input: "Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics.", }); console.log(response.output_text); ``` ```typescript Frozen configuration theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ model: "google/gemini-3.1-flash-lite", input: "Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics.", max_steps: 1, instructions: "", tools: [ { type: "web_search", search_context_size: "low", }, ], }); console.log(response.output_text); ``` ```bash Dynamic preset theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "fast-search", "input": "Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics." }' | jq ``` ```bash Frozen configuration theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "google/gemini-3.1-flash-lite", "input": "Explain the 2023 Nobel Prize in Physics: who won, what attosecond physics is, and why their work matters for studying electron dynamics.", "max_steps": 1, "instructions": "", "tools": [ { "type": "web_search", "search_context_size": "low" } ] }' | jq ``` ```json theme={null} { "id": "resp_a52a3c51-2a7a-4494-8dbf-1375f7d78613", "created_at": 1779895991, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "3 October 2023\nThe Royal Swedish Academy of Sciences has decided to award the Nobel Prize in Physics 2023 to\n**Pierre Agostini**\nThe Ohio State University, Columbus, USA\n**Ferenc Krausz**\nMax Planck Institute of Quantum Optics, Garching and Ludwig-Maximilians-Universität München, Germany\n**Anne L’Huillier**\nLund University, Sweden\n*“for experimental methods that generate attosecond pulses of light for the study of electron dynamics in matter”*\n...\nThe three Nobel Laureates in Physics 2023 are being recognised for their experiments, which have given humanity new tools for exploring the world of electrons inside atoms and molecules.\nPierre Agostini, Ferenc Krausz and Anne L’Huillier have demonstrated a way to create extremely short pulses of light that can be used to measure the rapid processes in which electrons move or change energy.\n...\nThe laureates’ experiments have produced pulses of light so short that they are measured in attoseconds, thus demonstrating that these pulses can be used to provide images of processes inside atoms and molecules.\n...\nIn 2001, **Pierre Agostini** succeeded in producing and investigating a series of consecutive light pulses, in which each pulse lasted just 250 attoseconds.\nAt the same time, **Ferenc Krausz** was working with another type of experiment, one that made it possible to isolate a single light pulse that lasted 650 attoseconds.", "title": "Press release: The Nobel Prize in Physics 2023 - NobelPrize.org", "url": "https://www.nobelprize.org/prizes/physics/2023/press-release/", "date": "2023-10-03", "last_updated": "2026-03-26", "source": "web" }, { "id": 2, "snippet": "*​Left to right: Pierre Agostini, Ferenc Krausz, Anne L’Huillier.\nArtwork by Niklas Elmehed.*\nThe Nobel Prize was given between these three people for their combined work in generating attosecond extreme ultraviolet (XUV) pulses through high harmonic generation.\nLaying the groundwork was Anne L’Huillier, who in 1987 discovered that XUV pulses could be produced by focusing a laser into a gas.\nFollowing this, in 2001 Pierre Agostini and Ferenc Krausz respectively proved that you could generate and measure XUV pulses that were only attoseconds long (one quintillionth of a second).\nAgostini showed that you could produce these pulses consistently, and Krausz showed that you could isolate single pulses.", "title": "CLF Nobel Prize in Physics 2023 for Attosecond Science", "url": "https://www.clf.stfc.ac.uk/Pages/Nobel-Prize-in-Physics-2023-for-Attosecond-Science.aspx", "date": "2023-10-16", "last_updated": "2026-03-26", "source": "web" }, { "id": 3, "snippet": "Therefore, the probability distribution, \\( |a_{\\textbf {p}}(\\tau )|^{2} \\), of finding an electron ionized in the continuum with momentum \\( {\\textbf {p}} \\), after the interaction has occurred (at \\( t=\\infty \\)), in a pump-probe experiments,\nwith an intense IR pulse and a delayed-attosecond XUV pulse, is given by:\n\\( a_{\\textbf {p}}(\\tau )=-i\\int _{-\\infty }^{\\infty }{\\textbf {E}}_{XUV}(t-\\tau )\\cdot {\\textbf {d}}[{\\textbf {p}}+{\\textbf {A}}_{IR}(t)]e^{+i(I_{P}t-S(t))}dt\\quad (1.7) \\)\nwith\n\\( S(t)={\\frac {1}{2}}|{\\textbf {p}}|^{2}t+\\int _{t}^{\\infty }({\\textbf {p}}\\cdot {\\textbf {A}}_{IR}(t')+{\\frac {1}{2}}|{\\textbf {A}}_{IR}(t')|^{2})dt' \\)\nEquation \\( (1.7) \\) describes the photoionization phenomenon of two-color interaction (XUV-IR) with a single-level atom and single active electron.", "title": "Attosecond physics - Wikipedia", "url": "https://en.wikipedia.org/wiki/Attosecond_physics", "date": "2005-09-15", "last_updated": "2026-05-12", "source": "web" }, { "id": 4, "snippet": "# Profile of Pierre Agostini, Anne L’ Huillier, and Ferenc Krausz: 2023 Nobel laureates in Physics\n...\nThe 2023 Nobel Prize in Physics was awarded to Pierre Agostini, Anne L’Huillier, and Ferenc Krausz for discoveries spanning many decades that led to the shortest possible pulses of electromagnetic radiation in time, the attosecond time domain (1 as = 10^18^ sec).\n...\nThe award of the 2023 Nobel Prize in Physics for attosecond pulses acknowledges, in a fundamental way, the limits of time itself, as well as a number of important strong-field atomic physics processes that allowed researchers to go beyond the pulse durations achievable with available lasers.", "title": "Profile of Pierre Agostini, Anne L' Huillier, and Ferenc Krausz - PMC", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC10835058/", "date": "2024-01-22", "last_updated": "2026-03-17", "source": "web" }, { "id": 5, "snippet": "Pierre Agostini, Ferenc Krausz and Anne L’Huillier have won the 2023 Nobel Prize for Physics for “experimental methods that generate attosecond pulses of light for the study of electron dynamics in matter”.\nL’Huillier is based at Sweden’s Lund University; Krausz is at the Max Planck Institute of Quantum Optics and the Ludwig-Maximillians University of Munich in Germany; and Agostini is at Ohio State University in the US.\nThe prize is worth 11 million kronor (one million dollars) and is split equally between the winners.\nThe prize will be presented in Stockholm on 10 December 2023.\n...\nL’Huillier, Agostini and Krausz pioneered the generation of ultrashort light pulses with durations in the attosecond (10^−18 ^s or 1 as) range.\nThis is roughly the time scale on which electrons in atoms, molecules and materials interact to give rise to chemical reactions and the properties of materials.\nAttosecond pulses allow researchers to observe these processes unfold, providing profound insights into the chemical and material world.", "title": "Pierre Agostini, Ferenc Krausz and Anne L'Huillier win 2023 Nobel ...", "url": "https://physicsworld.com/a/pierre-agostini-ferenc-krausz-and-anne-lhuillier-win-2023-nobel-prize-for-physics/", "date": "2023-10-03", "last_updated": "2026-05-22", "source": "web" } ], "type": "search_results", "queries": [ "2023 Nobel Prize in Physics attosecond pulses winners Agostini Krausz LHuillier", "definition of attosecond physics electron dynamics", "applications of attosecond physics to electron motion in atoms molecules materials" ] }, { "id": "msg_330acdd0-a511-4e74-bfe4-31875f6008ea", "content": [ { "text": "The 2023 Nobel Prize in Physics went to Pierre Agostini, Ferenc Krausz, and Anne L’Huillier for developing experimental methods to generate attosecond pulses of light, which make it possible to observe and control electron motion in matter in real time.[web:1][web:5] Their work opened a new time‑scale—attoseconds—where electron dynamics in atoms, molecules, and solids actually happen.[web:1][web:4][web:5]\n\n## Who won and for what\n\n- The laureates are **Pierre** Agostini (Ohio State University, USA), Ferenc Krausz (Max Planck Institute of Quantum Optics and LMU Munich, Germany), and Anne L’Huillier (Lund University, Sweden).[web:1][web:5] \n- The official citation is “for experimental methods that generate attosecond pulses of light for the study of electron dynamics in matter.”[web:1][web:5] \n- L’Huillier’s work, beginning in 1987, showed that focusing a strong laser into a gas creates a comb of high‑order harmonics in the extreme ultraviolet (XUV), laying the foundation for attosecond pulse generation.[web:1][web:2] \n\n## What attosecond physics is\n\n- An attosecond is 10⁻¹⁸ seconds; there are as many attoseconds in one second as there are seconds in about 31 billion years.[web:1][web:4][web:5] \n- Attosecond physics uses ultrashort XUV light pulses of this duration to probe and manipulate processes that occur on the natural time‑scale of electrons—such as ionization, excitation, and charge migration.[web:3][web:4][web:5] \n- In 2001, Agostini produced and characterized a train of light pulses where each pulse lasted about 250 attoseconds, while Krausz isolated a single pulse of about 650 attoseconds, demonstrating controlled access to this time regime.[web:1][web:2][web:4] \n\n## Why this matters for electron dynamics\n\n- Electrons move and rearrange on attosecond time‑scales, so conventional ultrafast lasers (femtoseconds, 10⁻¹⁵ s) are too slow to resolve the **details** of their motion; attosecond pulses act like a strobe light that can “freeze” and time‑tag these changes.[web:1][web:4][web:5] \n- With attosecond pump–probe experiments, researchers can watch how an electron leaves an atom (photoemission delays), how charge flows across a molecule, or how electrons respond in a solid during the very first instants of electrical conduction or light absorption.[web:3][web:4][web:5] \n- This capability enables deeper understanding and potential control of key phenomena such as chemical bond breaking and formation, ultrafast phase transitions in materials, and the speed limits of electronic and photonic devices.[web:4][web:5] \n\n## Examples of applications\n\n- In atoms and molecules, attosecond spectroscopy can measure time delays between electrons escaping from different orbitals, mapping how electron clouds reorganize during ionization and chemical reactions.[web:3][web:4][web:5] \n- In condensed matter, attosecond techniques can follow how electrons in a solid screen electric fields, carry currents, or transfer energy, informing the design of faster electronics, optoelectronic components, and novel quantum materials.[web:4][web:5] \n- More broadly, the work defines the practical limit of time resolution in light‑based measurements, giving science a tool to directly interrogate and eventually steer electron behavior rather than inferring it only from before‑and‑after states.[web:1][web:4][web:5]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 5536, "output_tokens": 886, "total_tokens": 6422, "cost": { "currency": "USD", "input_cost": 0.00244, "output_cost": 0.00886, "total_cost": 0.01675, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779895991, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Wednesday, May 27, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` ### pro-search Researched answers with tool use for most queries. ```python Dynamic preset theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( preset="pro-search", input="Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", ) print(response.output_text) ``` ```python Frozen configuration theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( model="openai/gpt-5.1", input="Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", max_steps=3, instructions="", tools=[ { "type": "web_search", "search_context_size": "medium", }, {"type": "fetch_url"}, ], ) print(response.output_text) ``` ```typescript Dynamic preset theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ preset: "pro-search", input: "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", }); console.log(response.output_text); ``` ```typescript Frozen configuration theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ model: "openai/gpt-5.1", input: "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", max_steps: 3, instructions: "", tools: [ { type: "web_search", search_context_size: "medium", }, { type: "fetch_url" }, ], }); console.log(response.output_text); ``` ```bash Dynamic preset theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "pro-search", "input": "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP." }' | jq ``` ```bash Frozen configuration theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "openai/gpt-5.1", "input": "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", "max_steps": 3, "instructions": "", "tools": [ { "type": "web_search", "search_context_size": "medium" }, {"type": "fetch_url"} ] }' | jq ``` ```json theme={null} { "id": "resp_e80c520d-2788-4b3d-8662-6a0e3d611e8f", "created_at": 1779391438, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "Abstract:The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration.\nThe best performing models also connect the encoder and decoder through an attention mechanism.\nWe propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.\nExperiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.\nOur model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU.\nOn the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature.\nWe show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.", "title": "[1706.03762] Attention Is All You Need - arXiv", "url": "https://arxiv.org/abs/1706.03762", "date": "2017-06-12", "last_updated": "2026-05-19", "source": "web" }, { "id": 2, "snippet": "At each layer, each token is then contextualized within the scope of the context window with other (unmasked) tokens via a parallel multi-head attention mechanism, allowing the signal for key tokens to be amplified and less important tokens to be diminished.\n...\nThe original version of the transformer architecture was proposed in the 2017 paper \"Attention Is All You Need\" by researchers at Google.\nThe predecessors of transformers were developed as an improvement over previous architectures for machine translation, but have found many applications since.\n...\nOn 2017-06-12, the original (100M-parameter) encoder–decoder transformer model was published in the \"Attention is all you need\" paper.\nAt the time, the focus of the research was on improving seq2seq for machine translation, by removing its recurrence to process all tokens in parallel, but preserving its dot-product attention mechanism to keep its text processing performance.\nThis led to the introduction of a multi-head attention model that was easier to parallelize due to the use of independent heads and the lack of recurrence.\nIts parallelizability was an important factor to its widespread use in large neural networks.", "title": "Transformer (deep learning) - Wikipedia", "url": "https://en.wikipedia.org/wiki/Transformer_(deep_learning)", "date": "2019-08-25", "last_updated": "2026-05-15", "source": "web" }, { "id": 3, "snippet": "We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train.", "title": "Attention is All you Need - NIPS papers", "url": "https://papers.nips.cc/paper/7181-attention-is-all-you-need", "date": "2017-01-01", "last_updated": "2026-05-21", "source": "web" }, { "id": 4, "snippet": "The paper introduced a new deep learning architecture known as the transformer, based on the attention mechanism proposed in 2014 by Bahdanau *et al.* The transformer approach it describes has become the main architecture of a wide variety of artificial intelligence, including large language models.\nAt the time, the focus of the research was on improving Seq2seq techniques for machine translation, but the authors go further in the paper, foreseeing the technique's potential for other tasks like question answering and what is now known as multimodal generative AI.\n...\nThe paper is best known for introducing the Transformer architecture, which underlies most modern large language models (LLMs).\nA key reason why the architecture is preferred by most modern LLMs is the parallelizability of the architecture over its predecessors.\nThis ensures that the operations necessary for training can be accelerated on a GPU, allowing both faster training times and models of bigger sizes to be trained.\n...\nOn 2017-06-12, the original (100M-parameter) encoder–decoder transformer model was published in the \"Attention is all you need\" paper.", "title": "Attention Is All You Need - Wikipedia", "url": "https://en.wikipedia.org/wiki/Attention_Is_All_You_Need", "date": "2023-12-04", "last_updated": "2026-05-17", "source": "web" }, { "id": 5, "snippet": "In this video, YC's Ankit Gupta traces how AI learned to understand language — from early RNNs and LSTMs to attention mechanisms and the breakthrough 2017 paper Attention Is All You Need — the discovery that unlocked the modern AI era.\n...\n{ts:17} [Music] A transformer is a neural network that uses self attention to take input data\n{ts:26} like text or images, model the relationships between that data, and finally generate outputs like meaningful\n{ts:32} text responses, translations, or classifications.\n...\nThen came the big breakthrough in 2017 when a team of researchers at\n{ts:387} Google published a paper called attention is all you need, which proposed a new machine translation\n{ts:392} architecture that they called a transformer.\n...\n{ts:409} version of the encoder decoder architecture originally proposed in seek to seek.\nInstead of compressing inputs\n{ts:414} into a single vector embedding, transformers kept separate embeddings for each input token and updated these\n{ts:420} through self attention, a mechanism that updated token representations based on a learned weighted dotproduct over the\n{ts:426} embeddings of all other tokens in the sequence.\nBecause each token in this architecture could attend to all others\n{ts:431} simultaneously, transformers could process an entire sequence in parallel, making them dramatically faster than\n{ts:437} RNN's.\nRemarkably, they were also much more accurate on machine translation benchmarks.", "title": "Transformers Explained: The Discovery That Changed AI Forever", "url": "https://www.youtube.com/watch?v=JZLZQVmfGn8&vl=en-US", "date": "2025-10-23", "last_updated": "2026-03-27", "source": "web" }, { "id": 6, "snippet": "●Presents a new neural architecture named the Transformer\n●Based solely on the attention mechanism widely used in SEQ2SEQ models\n●More parallelizable compared to existing state-of-the-art (SOTA) models\n●Achieves SOTA in 2 machine translation datasets\n…\n...\n• Less total computational complexity per layer\n• More parallelizable than existing fully autoregressive models\n• Shorten the path between tokens to enable model to learn long-term \ndependency better\n...\nIntroduces a groundbreaking new model that is solely based on attention\n●\nFaster and better than existing models\n●", "title": "[PDF] Attention Is All You Need", "url": "https://ysu1989.github.io/courses/au20/cse5539/Transformer.pdf", "date": null, "last_updated": "2026-05-15", "source": "web" }, { "id": 7, "snippet": "The Transformer architecture has revolutionized natural language processing (NLP) since its introduction, establishing itself as a cornerstone for modern advancements in the field.\nThis architecture offers significant improvements in handling a wide range of NLP tasks, from translation to text summarization, by overcoming limitations inherent in previous models.\nIntroduced by Vaswani et al. in the paper “Attention is All You Need” in 2017, the Transformer architecture is characterized by its use of self-attention mechanisms and feed-forward neural networks.\nUnlike its predecessors, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), which process data sequentially, the Transformer processes input data in parallel.\nThis parallelization enables the model to leverage modern hardware more efficiently, resulting in faster training times and improved scalability.\nAt the heart of the Transformer is the self-attention mechanism, which allows the model to weigh the significance of different words in a sentence relative to each other.\nThis feature is critical for understanding contextual relationships and dependencies, regardless of the distance between words in the input sequence.\n...\nThis capacity to capture global dependencies without the constraints of sequential data processing is a notable advantage over traditional models.\n...\nThe Transformer architecture’s flexibility makes it suitable for a broad spectrum of NLP applications.\nIn machine translation, it has significantly improved the accuracy and fluency of translated text.\nFor tasks like text summarization, sentiment analysis, and question answering, Transformers provide state-of-the-art results by effectively capturing and modeling complex language patterns.\n...\nIn summary, the Transformer architecture has fundamentally changed the landscape of NLP by introducing a novel approach to processing and understanding language data.\nIts innovative use of self-attention and parallel processing has paved the way for numerous breakthroughs, making it an essential tool for anyone looking to advance in the field of natural language processing.", "title": "What is the Transformer architecture in NLP? - Milvus", "url": "https://milvus.io/ai-quick-reference/what-is-the-transformer-architecture-in-nlp", "date": "2026-03-26", "last_updated": "2026-05-13", "source": "web" }, { "id": 8, "snippet": "In “Attention Is All You Need”, we introduce the Transformer, a novel neural network architecture based on a self-attention mechanism that we believe to be particularly well suited for language understanding.\nIn our paper, we show that the Transformer outperforms both recurrent and convolutional models on academic English to German and English to French translation benchmarks.\nOn top of higher translation quality, the Transformer requires less computation to train and is a much better fit for modern machine learning hardware, speeding up training by up to an order of magnitude.\n...\nIn contrast, the Transformer only performs a small, constant number of steps (chosen empirically).\nIn each step, it applies a self-attention mechanism which directly models relationships between all words in a sentence, regardless of their respective position.", "title": "Transformer: A Novel Neural Network Architecture for Language ...", "url": "https://research.google/blog/transformer-a-novel-neural-network-architecture-for-language-understanding/", "date": null, "last_updated": "2026-05-17", "source": "web" }, { "id": 9, "snippet": "A new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely is proposed, which generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.", "title": "[PDF] Attention is All you Need | Semantic Scholar", "url": "https://www.semanticscholar.org/paper/Attention-is-All-you-Need-Vaswani-Shazeer/204e3073870fae3d05bcbc2f6a8e263d9b72e776", "date": "2017-06-12", "last_updated": "2025-03-22", "source": "web" }, { "id": 10, "snippet": "Abstract:Natural Language Processing (NLP) has witnessed a transformative leap with the advent of transformer-based architectures, which have significantly enhanced the ability of machines to understand and generate human-like text.\nThis paper explores the advancements in transformer models, such as BERT and GPT, focusing on their superior performance in text understanding tasks compared to traditional methods like recurrent neural networks (RNNs).\nBy analyzing statistical properties through visual representations-including probability density functions of text length distributions and feature space classifications-the study highlights the models' proficiency in handling long-range dependencies, adapting to conditional shifts, and extracting features for classification, even with overlapping classes.\nDrawing on recent 2024 research, including enhancements in multi-hop knowledge graph reasoning and context-aware chat interactions, the paper outlines a methodology involving data preparation, model selection, pretraining, fine-tuning, and evaluation.\nThe results demonstrate state-of-the-art performance on benchmarks like GLUE and SQuAD, with F1 scores exceeding 90%, though challenges such as high computational costs persist.\nThis work underscores the pivotal role of transformers in modern NLP and suggests future directions, including efficiency optimization and multimodal integration, to further advance language-based AI systems.", "title": "Exploring Transformer-Based Architectures for Text Understanding", "url": "https://arxiv.org/abs/2503.20227", "date": "2025-03-26", "last_updated": "2026-02-21", "source": "web" } ], "type": "search_results", "queries": [ "Attention Is All You Need transformer paper core contributions", "why transformer architecture changed NLP", "Vaswani et al 2017 attention is all you need key results" ] }, { "id": "msg_b9dc751b-50fd-4941-ad3d-00bb2eb98f73", "content": [ { "text": "The paper “Attention Is All You Need” introduced the **Transformer**, a sequence‑to‑sequence model built entirely from attention and feed‑forward layers, and showed it could beat recurrent and convolutional models on machine translation while training much faster and scaling better.[web:1][web:8] This architectural shift made self‑attention the default building block for modern NLP and enabled today’s large language models.[web:2][web:4][web:7]\n\n## Core technical ideas\n\n- The Transformer removes recurrence and convolutions, using stacked self‑attention and position‑wise feed‑forward layers in both encoder and decoder.[web:1][web:3] Positional encodings are added to token embeddings so the model can represent order despite having no recurrence.[web:1] \n- It introduces **multi‑head** attention: several parallel attention heads project queries/keys/values, attend in different subspaces, then concatenate and linearly mix their outputs.[web:1][web:2] This lets the model capture different types of relations (e.g., syntax and coreference) in parallel.[web:1] \n- The paper emphasizes three attention types within the architecture: encoder self‑attention, decoder masked self‑attention (preventing access to future tokens), and encoder‑decoder attention where the decoder attends over encoder outputs.[web:1][web:8]\n\n## Empirical results in the paper\n\n- On WMT 2014 English→German translation, the base Transformer reaches 28.4 BLEU, surpassing prior state of the art, including ensembles, by over 2 BLEU.[web:1][web:3] On WMT 2014 English→French, a single Transformer achieves 41.8 BLEU after 3.5 days of training on 8 GPUs, outperforming earlier, much more expensive systems.[web:1] \n- The authors also show the model generalizes beyond translation by achieving strong results on English constituency parsing, even with limited labeled data.[web:1][web:9] These results demonstrate that attention‑only architectures are not just viable but competitive across sequence tasks.[web:1]\n\n## Why it was faster and more scalable\n\n- Self‑attention lets the model process all tokens in a sequence in parallel, unlike RNNs/LSTMs that must step through tokens sequentially, which makes it far more compatible with GPU/TPU hardware.[web:1][web:7] This parallelism reduced training time “by up to an order of magnitude” relative to recurrent models on the same tasks.[web:8] \n- For typical sequence lengths, the attention‑based layers in the paper have lower or comparable per‑layer computational complexity than recurrent or convolutional alternatives and much shorter paths between distant tokens, which helps learn long‑range dependencies.[web:1][web:6] This combination of efficiency and representational power is what made large‑scale pretraining feasible.[web:7][web:10]\n\n## How it changed NLP\n\n- The Transformer architecture became the foundation for major pretrained models such as BERT (encoder‑only), GPT (decoder‑only), and many others, which now dominate benchmarks in translation, question answering, summarization, and general language modeling.[web:2][web:7][web:10] These models exploit the same self‑attention blocks introduced in the 2017 paper, often with minor variants.[web:2] \n- By showing that attention‑only, highly parallel architectures can outperform traditional sequence models, the paper shifted the field’s default design pattern from RNNs/CNNs to transformers and opened the path to scaling parameters and data to today’s very large language models.[web:2][web:4][web:7] Its influence now extends beyond NLP to vision, speech, and multimodal models that reuse the same core transformer building blocks.[web:2][web:4]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 6620, "output_tokens": 863, "total_tokens": 7483, "cost": { "currency": "USD", "input_cost": 0.0038, "output_cost": 0.00863, "total_cost": 0.01788, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391438, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` ### deep-research In-depth analysis requiring multi-step reasoning. ```python Dynamic preset theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( preset="deep-research", input="What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?", ) print(response.output_text) ``` ```python Frozen configuration theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( model="openai/gpt-5.2", input="What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?", max_steps=10, instructions="", tools=[ { "type": "web_search", "search_context_size": "high", }, {"type": "fetch_url"}, ], ) print(response.output_text) ``` ```typescript Dynamic preset theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ preset: "deep-research", input: "What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?", }); console.log(response.output_text); ``` ```typescript Frozen configuration theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ model: "openai/gpt-5.2", input: "What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?", max_steps: 10, instructions: "", tools: [ { type: "web_search", search_context_size: "high", }, { type: "fetch_url" }, ], }); console.log(response.output_text); ``` ```bash Dynamic preset theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "deep-research", "input": "What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?" }' | jq ``` ```bash Frozen configuration theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "openai/gpt-5.2", "input": "What is the EU AI Act: its risk-based classification system, the prohibited-AI categories, and the general structure of obligations for high-risk AI systems?", "max_steps": 10, "instructions": "", "tools": [ { "type": "web_search", "search_context_size": "high" }, {"type": "fetch_url"} ] }' | jq ``` ```json theme={null} { "id": "resp_96f780f3-aeaf-4c6c-a86a-0c07dd49e057", "created_at": 1779391925, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "The AI Act sets out a risk-based rules for AI developers and deployers regarding specific uses of AI.\n...\nThe AI Act defines 4 levels of risk for AI systems:\n...\nAll AI systems considered a clear threat to the safety, livelihoods and rights of people are banned.\nThe **AI Act prohibits eight practices**, namely:\n1. harmful AI-based manipulation and deception\n2. harmful AI-based exploitation of vulnerabilities\n3. social scoring\n4. Individual criminal offence risk assessment or prediction\n5. untargeted scraping of the internet or CCTV material to create or expand facial recognition databases\n6. emotion recognition in workplaces and education institutions\n7. biometric categorisation to deduce certain protected characteristics\n8. real-time remote biometric identification for law enforcement purposes in publicly accessible spaces\nThe prohibitions became effective in February 2025.", "title": "AI Act | Shaping Europe's digital future - European Union", "url": "https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai", "date": "2026-04-09", "last_updated": "2026-05-20", "source": "web" }, { "id": 2, "snippet": "1. Irrespective of whether an AI system is placed on the market or put into service independently of the products referred to in points (a) and (b), that AI system shall be considered to be high-risk where both of the following conditions are fulfilled:\n(a) the AI system is intended to be used as a safety component of a product, or the AI system is itself a product, covered by the Union harmonisation legislation listed in Annex I;\n(b) the product whose safety component pursuant to point (a) is the AI system, or the AI system itself as a product, is required to undergo a third-party conformity assessment, with a view to the placing on the market or the putting into service of that product pursuant to the Union harmonisation legislation listed in Annex I.\n...\n3. By derogation from paragraph 2, an AI system referred to in Annex III shall not be considered to be high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making.\nThe first subparagraph shall apply where any of the following conditions is fulfilled:\n(a) the AI system is intended to perform a narrow procedural task;\n(b) the AI system is intended to improve the result of a previously completed human activity;\n(c) the AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review; or\n(d) the AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III.\nNotwithstanding the first subparagraph, an AI system referred to in Annex III shall always be considered to be high-risk where the AI system performs profiling of natural persons.", "title": "Article 6: Classification Rules for High-Risk AI Systems - EU AI Act", "url": "https://artificialintelligenceact.eu/article/6/", "date": null, "last_updated": "2026-05-20", "source": "web" }, { "id": 3, "snippet": "(a) remote biometric identification systems.\nThis shall not include AI systems intended to be used for biometric verification the sole purpose of which is to confirm that a specific natural person is the person he or she claims to be;\n(b) AI systems intended to be used for biometric categorisation, according to sensitive or protected attributes or characteristics based on the inference of those attributes or characteristics;\n(c) AI systems intended to be used for emotion recognition.", "title": "Annex III: High-Risk AI Systems Referred to in Article 6(2) - EU AI Act", "url": "https://artificialintelligenceact.eu/annex/3/", "date": null, "last_updated": "2026-05-17", "source": "web" }, { "id": 4, "snippet": "The EU AI Act introduces a proportionate risk-based approach to AI regulation, which imposes a gradual scheme of requirements and obligations depending on the level of risk posed to health, safety and fundamental rights.\n...\nIn line with this risk-based attitude, the final text of the Act has shifted from the binary low-risk vs. high-risk distinction proposed in the Commission’s White Paper on AI to a four-tiered risk framework, which classifies risk into four categories:\n- ‘unacceptable risks’ that lead to prohibited practices;\n- ‘high risks’ that trigger a set of detailed, complex and stringent obligations\n...\nBased on this risk classification scheme, certain AI practices are considered to entail unacceptable risks and directly prohibited in Article 5.\nThe exhaustive list of these prohibited AI practices includes use cases such as for social scoring and cognitive behavioral manipulation that may cause harm.\nHefty penalties are prescribed for carrying out these banned AI activities.", "title": "Key Issue 3: Risk-Based Approach - EU AI Act", "url": "https://www.euaiact.com/key-issue/3", "date": null, "last_updated": "2026-04-15", "source": "web" }, { "id": 5, "snippet": "As explained in our previous blog posts, the AI Act’s overall risk-based approach means that, depending on the level of risk, different requirements apply.\nIn total, there are four levels of risk:\n(1) unacceptable, in which case AI systems are prohibited (see our blog post);\n(2) high risk, in which case AI systems are subject to extensive requirements, including regarding transparency;\n(3) limited risk, which triggers only transparency requirements (see our blog post); and\n(4) minimal risk, which does not trigger any obligations.", "title": "What Are High-Risk AI Systems Within the Meaning of the EU's AI ...", "url": "https://www.wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/20240717-what-are-highrisk-ai-systems-within-the-meaning-of-the-eus-ai-act-and-what-requirements-apply-to-them", "date": "2024-07-17", "last_updated": "2026-05-19", "source": "web" }, { "id": 6, "snippet": "High-risk AI systems pursuant to Article 6(2) are the AI systems listed in any of the following areas:\n1.Biometrics, in so far as their use is permitted under relevant Union or national law:\n(a)remote biometric identification systems.This shall not include AI systems intended to be used for biometric verification the sole purpose of which is to confirm that a specific natural person is the person he or she claims to be;\n(b)AI systems intended to be used for biometric categorisation, according to sensitive or protected attributes or characteristics based on the inference of those attributes or characteristics;\n(c)AI systems intended to be used for emotion recognition.\n2.Critical infrastructure: AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity.\n3.Education and vocational training:", "title": "AI Act Service Desk - Annex III - European Union", "url": "https://ai-act-service-desk.ec.europa.eu/en/ai-act/annex-3", "date": null, "last_updated": "2026-05-17", "source": "web" }, { "id": 7, "snippet": "- Unacceptable risk is prohibited (e.g. social scoring systems and manipulative AI).\n...\n- deploying **subliminal, manipulative, or deceptive techniques** to distort behaviour and impair informed decision-making, causing significant harm.\n- **exploiting vulnerabilities** related to age, disability, or socio-economic circumstances to distort behaviour, causing significant harm.\n- **biometric categorisation systems** inferring sensitive attributes (race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation), except labelling or filtering of lawfully acquired biometric datasets or when law enforcement categorises biometric data.\n- **social scoring**, i.e., evaluating or classifying individuals or groups based on social behaviour or personal traits, causing detrimental or unfavourable treatment of those people.\n- **assessing the risk of an individual committing criminal offenses** solely based on profiling or personality traits, except when used to augment human assessments based on objective, verifiable facts directly linked to criminal activity.\n- **compiling facial recognition databases** by untargeted scraping of facial images from the internet or CCTV footage.\n- **inferring emotions in workplaces or educational institutions**, except for medical or safety reasons.\n- **‘real-time’ remote biometric identification (RBI) in publicly accessible spaces for law enforcement**, except when: - searching for missing persons, abduction victims, and people who have been human trafficked or sexually exploited;\n- preventing substantial and imminent threat to life, or foreseeable terrorist attack; or\n- identifying suspects in serious crimes (e.g., murder, rape, armed robbery, narcotic and illegal weapons trafficking, organised crime, and environmental crime, etc.).\n...\n- AI systems listed under Annex III are always considered high-risk if it profiles individuals, i.e. automated processing of personal data to assess various aspects of a person’s life, such as work performance, economic situation, health, preferences, interests, reliability, behaviour, location or movement.\n...\nHigh risk AI providers must:\n- Establish a **risk management system** throughout the high risk AI system’s lifecycle;\n- Conduct **data governance**, ensuring that training, validation and testing datasets are relevant, sufficiently representative and, to the best extent possible, free of errors and complete according to the intended purpose.\n- Draw up **technical documentation** to demonstrate compliance and provide authorities with the information to assess that compliance.\n- Design their high risk AI system for **record-keeping** to enable it to automatically record events relevant for identifying national level risks and substantial modifications throughout the system’s lifecycle.\n- Provide **instructions for use** to downstream deployers to enable the latter’s compliance.\n- Design their high risk AI system to allow deployers to implement **human oversight**.\n...\nIn addition to the four obligations above, providers of GPAI models with systemic risk must also:\n- Perform **model evaluations**, including conducting and documenting**adversarial testing** to identify and mitigate systemic risk.\n- **Assess and mitigate possible systemic risks**, including their sources.\n- **Track, document and report serious incidents** and possible corrective measures to the AI Office and relevant national competent authorities without undue delay.\n- Ensure an adequate level of **cybersecurity protection**.", "title": "High-level summary of the AI Act | EU Artificial Intelligence Act", "url": "https://artificialintelligenceact.eu/high-level-summary/", "date": "2024-02-27", "last_updated": "2026-05-19", "source": "web" }, { "id": 8, "snippet": "The act takes a tiered approach to the risks AI poses, applying different compliance requirements based on the type of risk each category of AI represents.\n...\nFor instance, entities that violate prohibitions on banned AI uses can be fined up to 40M Euro of 7% of turnover.\n...\n1. **Unacceptable Risk:** Systems in this category are banned outright due to the high-risk of violating fundamental rights.\n2. **High Risk: ** These systems are deemed to present high-risk as they could negatively impact the rights or safety of individuals.\nThese systems are subject to stringent compliance measures.\n3. **Limited Risk: ** These kinds of systems present lower levels of risk than high-risk systems but are still subject to transparency requirements.\nIndividuals who interact with AI systems in this category must be able to clearly understand when they are interacting with an AI system.\n...\nAnnex III of the EU AI Act specifies the major AI uses that should be classified as high-risk.\nThese include:\n- AI systems covered by existing industry sector or product safety regulations such as those for medical devices, vehicles, and toys.‍\n- AI systems involved in non-banned uses of biometric identification, or emotion recognition.\n- AI systems used to manage critical infrastructure, like energy grids or transportation systems‍.\n...\n1. **Unacceptable Risks:** AI practices that pose a clear threat to individuals' safety or fundamental rights are prohibited.\nThis includes systems designed for subliminal manipulation or those exploiting vulnerabilities of specific groups.", "title": "EU AI Act: Summary & Compliance Requirements - ModelOp", "url": "https://www.modelop.com/ai-governance/ai-regulations-standards/eu-ai-act", "date": null, "last_updated": "2026-05-19", "source": "web" }, { "id": 9, "snippet": "Article 5 of the AI Act essentially prohibits AI practices that materially distort peoples’ behavior or that raise serious concerns in democratic societies.\nAs explained in our previous blog post, this is part of the overall risk-based approach taken by the AI Act, which means that different requirements apply in accordance with the level of risk.\nIn total, there are four levels of risk: unacceptable, in which case AI systems are prohibited; high risk, in which case AI systems are subject to extensive requirements; limited risk, which triggers only transparency requirements; and minimal risk, which does not trigger any obligations.\n...\nThe AI Act prohibits placing AI systems on the European Union’s market, putting them into service, or using them in the European Union to materially distort people’s behavior in a manner that causes or is likely to cause them physical or psychological harm:\n...\n- **Prohibited Systems.** The AI Act prohibits placing on the market, putting into service and using the following AI systems: - **Subliminal, manipulative and deceptive systems.** AI systems that deploy subliminal techniques beyond a person’s consciousness or purposefully use manipulative or deceptive techniques that materially distort people’s behavior by appreciably impairing their ability to make informed decisions.\nSuch systems cause people to make decisions that they would not have otherwise taken, [likely] resulting in significant harm.\n- **Exploiting vulnerabilities.** AI systems that exploit people’s vulnerabilities due to their age, disability, or social or economic situation.\nSuch systems also distort people’s behavior, [likely] resulting in significant harm.\n- **Facial recognition databases.** AI systems that create or expand facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage.\n- **Inferring emotions.\n** AI systems that infer emotions of individuals in the areas of workplace and educational institutions, except for AI medical or safety systems.\n- **Biometric categorization.** AI systems that categorize individual natural persons based on their biometric data to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life, or sexual orientation.\n...\nThe AI Act prohibits placing AI systems on the European Union’s market, putting them into service or using them in the European Union for social scoring or “minority report” scenario purposes.", "title": "Prohibited AI Practices—A Deep Dive into Article 5 of the European ...", "url": "https://www.wilmerhale.com/en/insights/blogs/wilmerhale-privacy-and-cybersecurity-law/20240408-prohibited-ai-practices-a-deep-dive-into-article-5-of-the-european-unions-ai-act", "date": "2024-04-08", "last_updated": "2026-04-18", "source": "web" }, { "id": 10, "snippet": "The EU AI Act (“AI Act”)^[1] ^takes a risk based approach and divides AI systems into four different risk groups: (1) prohibited AI practices, (2) high risk, (3) limited risk and (4) minimal risk AI systems.\n...\nHigh-risk AI systems are those that pose a significant risk to health, safety, or fundamental rights and are specifically classified as high-risk in the AI Act.", "title": "EU AI Act: obligations for high-risk AI systems - A&O Shearman", "url": "https://www.aoshearman.com/en/insights/ao-shearman-on-tech/zooming-in-on-ai-10-eu-ai-act-what-are-the-obligations-for-high-risk-ai-systems", "date": "2024-10-29", "last_updated": "2026-05-20", "source": "web" }, { "id": 11, "snippet": "There are three ways in which the legislation provides for AI systems to be considered ‘high-risk’:\n- when the AI system is itself a certain type of product;\n- when the AI system is a safety component of a certain type of product;\n- when the AI system meets the description of listed ‘high-risk’ AI systems.\n...\n###### Biometrics##Some AI systems that involve the processing of biometric data are entirely prohibited under the EU AI Act.\nFor example, AI-based biometric categorisation systems that categorise people based on their biometric data to deduce or infer their race, political opinions, trade union membership, religious or philosophical beliefs, sex life or sexual orientation are banned – subject to some limited exceptions in the context of law enforcement.", "title": "A guide to high-risk AI systems under the EU AI Act - Pinsent Masons", "url": "https://www.pinsentmasons.com/out-law/guides/guide-to-high-risk-ai-systems-under-the-eu-ai-act", "date": "2024-02-13", "last_updated": "2026-05-21", "source": "web" } ], "type": "search_results", "queries": [ "EU AI Act risk-based approach prohibited practices", "EU Artificial Intelligence Act high-risk AI obligations structure", "text of EU AI Act Annex III high-risk systems" ] }, { "id": "msg_1ae3e916-257c-454f-b89d-acf5d14e9f91", "content": [ { "text": "The EU AI Act is a horizontal regulation that classifies AI systems into four risk tiers (minimal, limited, high, and unacceptable), with escalating obligations as risk increases and outright bans for “unacceptable risk” systems.[web:4][web:5] High‑risk systems face a detailed set of lifecycle obligations around risk management, data, technical documentation, logging, transparency, human oversight, robustness, and governance.[web:7][web:10]\n\n## Risk-based system\n\nThe Act organizes AI into four main **risk** levels, linked to the potential impact on health, safety, and fundamental rights.[web:4][web:5]\n\n- Unacceptable risk: AI practices listed in Article 5 are prohibited and cannot be placed on the EU market, put into service, or used.[web:4][web:9] \n- High risk: AI systems that pose significant risks (especially in safety-critical and rights-sensitive domains) and are listed in Annex I/III or meet Article 6 criteria.[web:2][web:6] \n- Limited risk: AI systems subject mainly to specific transparency obligations (for instance, making clear that users are interacting with AI or are exposed to emotion recognition or deepfakes).[web:4][web:5] \n- Minimal risk: The vast majority of AI applications; they can be used freely without specific obligations under the AI Act beyond existing law.[web:4][web:5]\n\n## Prohibited AI categories (“unacceptable risk”)\n\nArticle 5 contains an exhaustive list of AI practices deemed an unacceptable threat to fundamental rights or democratic values.[web:4][web:9] These practices are banned subject to narrow exceptions, mainly in law enforcement contexts.[web:1][web:7]\n\nKey prohibited categories include:[web:1][web:7][web:9] \n- Manipulative and deceptive systems: AI that uses subliminal techniques or purposefully manipulative/deceptive techniques to materially distort behavior and significantly impair informed decision‑making, causing or likely causing physical or psychological harm. \n- Exploitation of vulnerabilities: AI that exploits vulnerabilities of a specific group due to age, disability, or socio‑economic situation in a way that distorts behavior and causes or is likely to cause significant harm. \n- Social scoring: Systems that evaluate or classify individuals or groups based on social behavior or personal traits, leading to detrimental or unfavorable treatment that is unjustified or disproportionate. \n- Predictive policing of individuals: AI used to assess or predict an individual’s risk of committing criminal offenses solely on profiling or personality traits, except where it only supports human assessments based on objective, verifiable facts linked to crime. \n- Untargeted facial scraping: Use of AI to create or expand facial recognition databases through untargeted scraping of images from the internet or CCTV footage. \n- Biometric categorisation for sensitive traits: AI that categorises people based on biometric data to infer sensitive attributes (such as race, political opinions, religious beliefs, trade union membership, sex life, or sexual orientation), with narrow exceptions for certain law‑enforcement dataset operations.[web:7][web:9] \n- Emotion recognition at work and in education: AI systems that infer emotions of individuals in workplaces or educational institutions, except for medical or safety purposes.[web:1][web:7][web:9] \n- Real‑time remote biometric identification in public spaces for law enforcement: Broadly prohibited, but allowed under strict conditions for narrowly defined purposes (e.g. searching for missing persons, preventing imminent threats or terrorist attacks, or identifying suspects of serious crimes) with safeguards.[web:1][web:7]\n\n## How “high-risk” AI is defined\n\nHigh‑risk systems are defined both by their role in regulated products and by specific use cases listed in Annex III.[web:2][web:6] The classification is not purely sectoral; it is anchored in the potential significant risk to health, safety, or fundamental rights.[web:2][web:10]\n\nMain channels into the “high‑risk” category:[web:2][web:11] \n- Safety components or products under existing EU product safety law: AI that is a safety component of, or itself is, a product regulated under sectoral safety legislation listed in Annex I (e.g. certain machinery, toys, medical devices, vehicles), when that product requires third‑party conformity assessment. \n- Standalone systems in Annex III areas: Annex III lists specific high‑risk use cases, including: \n - Biometrics: remote biometric identification (non‑prohibited uses), certain biometric categorisation and emotion recognition uses.[web:3][web:6] \n - Critical infrastructure: safety components for managing road traffic, digital infrastructure, or supply of water, gas, heating, or electricity.[web:6] \n - Education and vocational training: systems influencing access to or evaluation within education and training.[web:6] \n - Employment and worker management: systems for recruitment, CV screening, promotion, and termination decisions.[web:6] \n - Access to essential services and benefits: e.g. systems used in creditworthiness, social benefits, healthcare triage.[web:6] \n - Law enforcement, migration, asylum and border control, and administration of justice: systems that materially influence decisions in these areas.[web:6]\n\nThe Act also contains a “de‑scoping” rule: some Annex III systems are not treated as high‑risk if they only perform narrow procedural or preparatory tasks, merely improve previous human activity, or detect patterns without materially influencing outcomes; however, profiling of natural persons in Annex III use cases remains high‑risk.[web:2][web:7]\n\n## Core obligations for high-risk AI systems\n\nHigh‑risk providers and deployers must meet extensive, prescriptive obligations across the system lifecycle, centered on safety, fundamental rights, and accountability.[web:7][web:10] Obligations primarily apply to “providers” (those who develop/place systems on the market), with additional duties for “deployers” (users).[web:7][web:10]\n\nKey obligations for providers of high‑risk systems include:[web:7][web:10] \n- Risk management system: Establish and maintain a documented, continuous risk management process covering design, testing, and post‑market monitoring, including identification, analysis, and mitigation of reasonably foreseeable risks.[web:7][web:10] \n- Data governance and quality: Ensure training, validation, and testing datasets are relevant, sufficiently representative, appropriately balanced where feasible, and as free of errors and bias as possible for the intended purpose, with documented data governance and data management practices.[web:7][web:10] \n- Technical documentation: Prepare and keep up‑to‑date technical documentation that demonstrates conformity with AI Act requirements and enables authorities to assess compliance (architecture, algorithms, data sources, performance metrics, risk controls, etc.).[web:7][web:10] \n- Logging and record‑keeping: Design systems with automatic logging to record events relevant to operation and risk, enabling traceability, incident analysis, and supervision.[web:7][web:10] \n- Transparency and instructions for use: Provide clear instructions and information for deployers (intended purpose, performance characteristics, limitations, data needs, human oversight measures, residual risks), and ensure that outputs are sufficiently understandable to users.[web:7][web:10] \n- Human oversight: Design systems so they can be effectively overseen by humans, including tools and interfaces allowing human operators to understand system functioning, detect anomalies, intervene, override, or stop the system where appropriate.[web:7][web:10] \n- Robustness, accuracy, and cybersecurity: Ensure systems achieve appropriate levels of accuracy, robustness, resilience to errors or attacks, and cybersecurity, commensurate with the system’s purpose and the risks involved.[web:7][web:10] \n- Quality management system and conformity assessment: Implement an internal quality management system and undergo conformity assessment procedures (often involving notified bodies) before placing the system on the market, then affix CE marking and keep post‑market monitoring and incident reporting in place.[web:7][web:10]\n\nDeployers of high‑risk systems also bear obligations, such as using systems in line with the provider’s instructions, performing context‑specific fundamental‑rights and risk assessments where required, ensuring qualified human oversight, and keeping logs and records from system use.[web:7][web:10]\n\n## Overall structure of obligations\n\nThe AI Act’s obligations scale with risk: prohibited practices are banned outright; high‑risk systems face full lifecycle compliance requirements; limited‑risk systems mainly require transparency; and minimal‑risk systems remain largely unregulated under this instrument.[web:4][web:5] Within high‑risk, obligations are structured around governance (risk management, QMS), technical controls (data, robustness, cybersecurity), operational controls (logging, human oversight), and documentation/transparency to regulators and users.[web:7][web:10]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 7607, "output_tokens": 1912, "total_tokens": 9519, "cost": { "currency": "USD", "input_cost": 0.00503, "output_cost": 0.01912, "total_cost": 0.0296, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391925, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` ### advanced-deep-research Institutional-grade research with maximum depth. ```python Dynamic preset theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( preset="advanced-deep-research", input="Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth.", ) print(response.output_text) ``` ```python Frozen configuration theme={null} from perplexity import Perplexity client = Perplexity() response = client.responses.create( model="anthropic/claude-opus-4-6", input="Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth.", max_steps=10, instructions="", tools=[ { "type": "web_search", "search_context_size": "high", }, {"type": "fetch_url"}, ], ) print(response.output_text) ``` ```typescript Dynamic preset theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ preset: "advanced-deep-research", input: "Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth.", }); console.log(response.output_text); ``` ```typescript Frozen configuration theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); const response = await client.responses.create({ model: "anthropic/claude-opus-4-6", input: "Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth.", max_steps: 10, instructions: "", tools: [ { type: "web_search", search_context_size: "high", }, { type: "fetch_url" }, ], }); console.log(response.output_text); ``` ```bash Dynamic preset theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "advanced-deep-research", "input": "Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth." }' | jq ``` ```bash Frozen configuration theme={null} curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "anthropic/claude-opus-4-6", "input": "Provide a competitive analysis of AWS, Azure, and Google Cloud across IaaS market share, pricing models for compute and storage, and AI/ML service depth.", "max_steps": 10, "instructions": "", "tools": [ { "type": "web_search", "search_context_size": "high" }, {"type": "fetch_url"} ] }' | jq ``` ```json theme={null} { "id": "resp_25a8fcf0-0475-46f9-b6b5-0b16fe8071d9", "created_at": 1779391483, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "As of Q2 2025,\n**Amazon Web Services (AWS) holds 30% of the global cloud infrastructure market**, maintaining its position as the dominant provider.\nMicrosoft Azure holds 20% and Google Cloud has 13%.\nTogether, The Big Three (AWS, Azure, and Google Cloud) control 63% of the rapidly growing cloud infrastructure market.- The global cloud infrastructure market reached **$99 billion in revenue** in Q2 2025\n...\n**Amazon Web Services (AWS)** is the market leader with 30% share, ahead of**Microsoft Azure** at 20% and**Google Cloud** at 13%.\nCombined, these \"Big Three\" providers control more than 60% of the worldwide cloud infrastructure market as of Q2 2025.\n...\n|Cloud Provider|Market Share (Q2 2025)|\n|--|--|\n|AWS|30%|\n|Microsoft Azure|20%|\n|Google Cloud|13%|\n|Alibaba Cloud|4%|\n...\nDespite strong competition from Microsoft and Google, AWS remains ahead of its closest competitor, Microsoft Azure, by 10 percentage points.\n**30% of the global cloud infrastructure market** as of Q2 2025.\n...\nMicrosoft Azure holds\n**20% of the global cloud infrastructure market** in Q2 2025, which makes it the second-largest provider.\nAzure trails AWS by 10 percentage points but maintains a significant lead over the third-place Google Cloud.\n...\nGoogle Cloud holds\n**13% of the global cloud infrastructure market** as of Q2 2025, securing third place among cloud providers.\nGoogle Cloud is the smallest of the \"Big Three\" but still commands a larger market share than all other competitors combined outside of AWS and Azure.\n...\nAll other cloud infrastructure providers combined account for less than 10% of the total market, illustrating the strong dominance of AWS, Azure, and Google Cloud.", "title": "AWS vs Azure vs Google: Cloud Market Share (2025) - Cargoson", "url": "https://www.cargoson.com/en/blog/global-cloud-infrastructure-market-share-aws-azure-google", "date": "2025-10-31", "last_updated": "2026-05-18", "source": "web" }, { "id": 2, "snippet": "Three main players dominate the machine learning platforms space:\n- **AWS** maintains overall cloud market leadership with approximately 30%\n- **Microsoft Azure** holds about 21% market share\n- **Google Cloud Platform** has grown to roughly 12% of the market\nThe market shares for AI-specific cloud workloads differ from general cloud services.\nAccording to IoT Analytics' research, AWS captures 34% of new public cloud AI projects, Microsoft Azure leads with 45%, and Google Cloud accounts for 17%.\nAll three provide comprehensive machine learning platforms with different strengths.", "title": "Best Cloud for AI Development 2025", "url": "https://www.aliz.ai/en/blog/comparing-machine-learning-platforms-finding-the-best-cloud-for-ai-development", "date": "2025-06-10", "last_updated": "2026-05-17", "source": "web" }, { "id": 3, "snippet": "After having established itself as an early leader in the market for cloud infrastructure, AWS is still ahead of the pack.\nAccording to estimates from Synergy Research Group, Amazon’s market share in the worldwide cloud infrastructure market amounted to 28 percent in the first quarter of 2026, ahead of Microsoft's Azure platform at 21 percent and Google Cloud at 14 percent.\nTogether, the \"Big Three\" hyperscalers account for more than 60 percent of the ever-growing cloud market, with the rest of the competition stuck in the low single digits.", "title": "Big Three Hold Dominant Lead in Accelerating Cloud Market - Statista", "url": "https://www.statista.com/chart/18819/worldwide-market-share-of-leading-cloud-infrastructure-service-providers/", "date": "2026-05-05", "last_updated": "2026-05-19", "source": "web" }, { "id": 4, "snippet": "While it has a smaller market share compared to AWS and Azure, Google Cloud is known for its strong performance in data analytics, Kubernetes, and high-performance networking.", "title": "AWS vs Azure vs Google: Cloud Services Comparison - Varonis", "url": "https://www.varonis.com/blog/aws-vs-azure-vs-google", "date": "2025-01-15", "last_updated": "2026-05-20", "source": "web" }, { "id": 5, "snippet": "AWS is the frontrunner in revenue, at $30.9 billion during Q2 2025.\nGoogle Cloud earned an impressive $13.6 billion during the same quarter.\nAnd Microsoft’s Intelligent Cloud group – with Azure as its primary service – grew dramatically in Q2 2025 (Microsoft’s fiscal Q4 2025), generating $29.9 billion in revenue.\n...\n## The State of Cloud Market Share in 2025\nAmazon (AMZN) is leading the way in cloud market share as we approach 2026.\nHere are the numbers based on Synergy Research Group data for Q2 2025:\n- **AWS**: 30%\n- **Azure**: 20%\n- **Google Cloud**: 13%\n...\nAzure’s revenue growth came in at an estimated 39% YoY.\nAnd Google Cloud’s revenue grew 32% YoY.\n...\nSo far, Google Cloud still trails behind AWS and Azure in market share.\nBut it’s demonstrating growth that’s impossible to ignore, especially in generative AI and analytics.\n...\nGoogle Cloud doesn’t have the market share leadership advantage like AWS.\nBut its YoY revenue growth beat AWS and Microsoft’s Intelligent Cloud group as a whole (though it trailed Azure’s growth rate).", "title": "Azure vs. AWS vs. Google Cloud: Who Wins the AI Cloud War?", "url": "https://marketwise.com/investing/azure-vs-aws-vs-google-cloud-whos-winning-the-cloud-ai-war-in-2025/", "date": "2025-10-10", "last_updated": "2026-05-21", "source": "web" }, { "id": 6, "snippet": "The global cloud market exploded in the third quarter of 2025 to $107 billion, increasing $7.6 billion in just one quarter—as AWS, Microsoft and Google Cloud continue to dominate global cloud market share.\nAWS, Google Cloud and Microsoft—combined—won 62 percent share of the global enterprise cloud infrastructure services market, according to new data from Synergy Research Group.\n...\n*CRN* breaks down the top five companies that won the most cloud market share in the booming infrastructure services market in the third quarter of 2025: AWS, Alibaba Cloud, Google Cloud, Microsoft and Oracle.\n...\nHere are the exact Q3 2025 cloud market share figures of AWS, Microsoft, Oracle, IBM and Google Cloud.\n...\nSoftware giant Oracle ranks No. 5 for share in the global cloud infrastructure services market by winning a 3 percent share during the third quarter of 2025.\n...\n**Q3 2025 Market Share: 13 Percent**\n...\nThe Mountain View, Calif.-based company won third place by winning 13 percent market share in the third quarter of 2025.\n...\nThe world’s largest software company won second place by winning a 20 percent share of the global market in Q3 2025.\n...\n**No. 1: Amazon Web Services**\n**Q3 2025 Market Share: 29 Percent**\n**Q2 2025 Market Share: 30 Percent**\n**Q3 2024 Market Share: 31 Percent**\nAWS continues to be the leading cloud company in the world by winning 29 percent share of the global market.", "title": "Global Cloud Market Share Q3 2025: AWS Lowers, Microsoft ... - CRN", "url": "https://www.crn.com/news/cloud/2025/global-cloud-market-share-q3-2025-aws-lowers-microsoft-and-google-stay-same", "date": "2025-11-06", "last_updated": "2026-05-13", "source": "web" }, { "id": 7, "snippet": "The results show that OCI is the cheapest option.\nAWS and Azure are competitively priced, while GCP is slightly more expensive.\nAs instances scale up, prices increase proportionally across all providers, with GCP consistently remaining the most expensive, albeit only marginally.\n...\nIn summary, for these specified storage amounts and regions, Azure is invariably cheaper, while AWS consistently charges higher prices.", "title": "Cloud pricing comparison - AWS, Azure, GCP, Oracle - EffectiveSoft", "url": "https://www.effectivesoft.com/blog/cloud-pricing-comparison.html", "date": "2025-01-16", "last_updated": "2026-05-20", "source": "web" }, { "id": 8, "snippet": "Among the hyperscalers, **Amazon Web Services (AWS)** maintains a **strong lead with a 30% global market share**, followed by **Microsoft Azure at 20%** and **Google Cloud at 13%**.\n...\n|Provider|Approx. Global Market Share (2025)|Key Notes|\n|--|--|--|\n|Amazon Web Services (AWS)|~30% in Q2 2025.|Still the leader, though share has dipped slightly from ~32-33%.|\n|Microsoft Azure|~20% in Q2 2025.|Second place, but growth remains strong.|\n|Google Cloud (GCP)|~13% in Q2 2025.|Gaining ground, especially via AI/data workloads.|\n|Alibaba Cloud|~4% globally.|Dominant in China, but global footprint remains limited.|\n|Oracle Cloud Infrastructure|~3% globally.|Strong niche in enterprise ERP + database tie-ins.|\n...\nIts global share (~20%) hides pockets of higher penetration in enterprise sectors and regions.\n**GCP**: Google’s strength in data analytics, machine learning, and AI is paying off — its share climbed to ~13% in Q2 2025.\nIt still trails in enterprise footing, but growing fast.\n...\nAs of 2026, **Amazon Web Services (AWS)** remains the global leader in cloud infrastructure, holding around **31% of the market share**, followed by **Microsoft Azure (25%)** and **Google Cloud Platform (11%)**.", "title": "Cloud Market share 2026: Top cloud providers and trends - Holori", "url": "https://holori.com/cloud-market-share-2026-top-cloud-vendors-in-2026/", "date": "2025-11-14", "last_updated": "2026-05-14", "source": "web" }, { "id": 9, "snippet": "- **There’s no universal “best.”** Match the platform to your **data location**, **team skills**, **regulatory constraints**, and **latency/throughput** goals.\n- **AWS SageMaker** = maximum flexibility and industrial-grade endpoint scaling — great once workloads stabilize.\n- **GCP Vertex AI** = fastest path from dataset to results (BigQuery + Vertex, TPUs) with sensible managed defaults.", "title": "Cloud AI Platforms: AWS vs. GCP vs. Azure for Machine-Learning ...", "url": "https://medium.com/@virtualik/cloud-ai-platforms-aws-vs-gcp-vs-azure-for-machine-learning-workloads-6033e162ce5c", "date": "2025-11-06", "last_updated": "2025-11-09", "source": "web" } ], "type": "search_results", "queries": [ "AWS vs Azure vs Google Cloud IaaS market share 2025", "AWS Azure Google Cloud pricing models compute storage comparison", "AWS Azure Google Cloud AI ML services comparison 2025" ] }, { "id": "msg_c8bc8d16-e6ef-4a6a-8d82-9c1288e68e6b", "content": [ { "text": "AWS leads IaaS market share globally, with Azure solidly second and Google Cloud third, while all three offer broadly similar pay‑as‑you‑go pricing models but differ in discounts and AI/ML strengths. [web:1][web:3][web:8] Azure and Google Cloud are often more cost‑competitive for specific compute and storage patterns, whereas AWS generally has the broadest service catalog and maturity, and Google leans hardest into integrated data + AI tooling. [web:4][web:7][web:9] \n\n## IaaS market share\n\n- Recent industry estimates for cloud infrastructure services put AWS at about 28–31% global share, Azure around 20–21%, and Google Cloud roughly 13–14%. [web:1][web:3][web:6][web:8] \n- Together, the three control more than 60% of the global infrastructure market, with AWS keeping a single‑digit lead over Azure and a much larger gap over Google Cloud. [web:1][web:3][web:5][web:8] \n\n**Implications**: \n- AWS: perceived safest “default” with the deepest ecosystem of third‑party tools and talent. [web:4][web:5][web:8] \n- Azure: particularly strong in Microsoft‑centric enterprises (Windows Server, SQL Server, M365, Active Directory). [web:4][web:5][web:8] \n- Google Cloud: smaller but growing quickly, especially for data, analytics, and AI‑heavy workloads. [web:4][web:5][web:8] \n\n## Compute pricing models\n\nAll three use on‑demand, discount, and commitment‑based models, but with different emphases. [web:7][web:9] \n\n- **AWS (EC2, Fargate, etc.)** \n - On‑demand per‑second or per‑hour pricing, plus Reserved Instances and Savings Plans for 1–3‑year commitments at significant discounts. [web:4][web:7] \n - Spot Instances give steep discounts for interruptible workloads, but AWS instance families and pricing options are the most complex to navigate. [web:4][web:7] \n\n- **Azure (Virtual Machines, Container Instances)** \n - Pay‑as‑you‑go, Reserved VM Instances, and Azure Savings Plans follow a structure similar to AWS, with hybrid benefit discounts for Windows and SQL licenses. [web:4][web:7] \n - Independent benchmarks frequently find Azure slightly cheaper than AWS for comparable VM configurations in common regions, especially when hybrid benefits apply. [web:7] \n\n- **Google Cloud (Compute Engine, Cloud Run)** \n - On‑demand with “sustained use” discounts that apply automatically as you run VMs longer, plus committed use discounts for 1–3 years. [web:4][web:7] \n - In comparisons using like‑for‑like instances, GCP often ends up marginally more expensive overall than AWS and Azure, though differences are workload‑ and region‑dependent. [web:7] \n\n**Competitive takeaways**: \n- For Linux‑heavy, steady workloads with long commitments, all three can be priced similarly once discounts are applied. [web:7] \n- Azure can be most cost‑efficient for Microsoft‑licensed stacks, AWS is most flexible but sometimes pricier, and Google favors automatic discounts that reduce planning overhead. [web:4][web:7] \n\n## Storage pricing models\n\nObject and block storage are structurally similar across providers, with differences in details, tiers, and egress. [web:4][web:7] \n\n- **AWS (S3, EBS, EFS)** \n - S3 uses tiered pricing (Standard, Infrequent Access, Glacier tiers) with separate charges for storage, API operations, and network egress. [web:4] \n - Comparative analyses often find AWS charging more than Azure for similar object and block storage footprints in the same regions. [web:7] \n\n- **Azure (Blob Storage, Managed Disks, Files)** \n - Azure Blob offers hot, cool, and archive tiers with competitive per‑GB rates, plus reserved capacity discounts. [web:4] \n - For common storage sizes and regions, Azure frequently comes out cheaper than both AWS and GCP in evaluated scenarios. [web:7] \n\n- **Google Cloud (Cloud Storage, Persistent Disks, Filestore)** \n - Multi‑regional, regional, nearline, coldline, and archive tiers, with pricing similar in structure to S3 and Blob Storage. [web:4] \n - Some cross‑provider benchmarks show GCP slightly more expensive overall, though the differences can be small and sensitive to egress and access patterns. [web:7] \n\n**Competitive takeaways**: \n- For storage‑heavy workloads, Azure often has a **cost** edge; AWS has the most mature ecosystem and features; GCP can be compelling when tightly integrating with BigQuery and analytics. [web:4][web:7] \n\n## AI/ML service depth\n\nAll three have full‑stack AI/ML platforms but with different strengths. [web:2][web:4][web:9] \n\n### AWS\n\n- Core services include Amazon SageMaker (end‑to‑end ML), Bedrock (managed foundation models and generative AI), and a wide range of task‑specific APIs for vision, speech, and personalization. [web:4][web:9] \n- In cloud AI project share, AWS holds a large portion but trails Azure for new AI projects in some analyses, reflecting strong general cloud leadership but stiffer competition in AI. [web:2] \n\n### Azure\n\n- Azure AI Studio and Azure Machine Learning integrate closely with Microsoft Fabric, Azure Databricks, and Power Platform, emphasizing enterprise governance and MLOps. [web:2][web:4] \n- Some industry research estimates Azure capturing the largest share of new public‑cloud AI projects, leveraging its enterprise relationships and integration with tools like GitHub Copilot and Microsoft 365. [web:2][web:5] \n\n### Google Cloud\n\n- Vertex AI provides managed training, tuning, and deployment, tightly integrated with BigQuery, plus access to Gemini and other foundation models. [web:2][web:4][web:9] \n- GCP is widely regarded as particularly strong for data analytics and ML workflows, aided by BigQuery, Dataflow, and TPUs, and is seeing growth in AI‑centric workloads despite smaller overall cloud share. [web:2][web:5][web:8][web:9] \n\n**Competitive takeaways**: \n- AWS: broadest **breadth** and industrial‑strength ML operations; good for large, mixed workloads that need tight integration with existing AWS services. [web:4][web:9] \n- Azure: strongest enterprise AI story, especially where Microsoft productivity tools and data platforms are already standard. [web:2][web:4][web:5] \n- Google Cloud: most opinionated and streamlined path for data‑to‑ML, appealing for analytics‑first, research, and ML‑heavy organizations. [web:2][web:4][web:9] \n\n## Strategic positioning summary\n\n- AWS is the overall IaaS leader with the deepest catalog, often at a pricing premium but with unmatched ecosystem maturity. [web:1][web:3][web:4][web:8] \n- Azure competes aggressively on price for compute and storage, integrates deeply with Microsoft software, and is very strong in enterprise and AI adoption. [web:5][web:7] \n- Google Cloud is smaller in IaaS share but highly competitive in data and AI, with a modern developer experience and growing share of AI‑driven workloads. [web:2][web:5][web:8][web:9] \n\nIf you share your primary workload patterns (e.g., Windows vs Linux, batch vs real‑time, AI intensity, data volumes), a more tailored recommendation across the three can be outlined.", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 6451, "output_tokens": 1774, "total_tokens": 8225, "cost": { "currency": "USD", "input_cost": 0.00358, "output_cost": 0.01774, "total_cost": 0.02677, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391483, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` ## Customizing Presets Presets provide sensible defaults, but you can override any parameter by passing additional parameters alongside the preset. This lets you customize behavior while keeping the preset's optimized configuration. You can always override the preset's search configuration by passing `search_context_size: "low"`, `"medium"`, or `"high"` — or use explicit `max_tokens` and `max_tokens_per_page` values for exact budget control. See [Configuring Search](/docs/agent-api/tools/web-search#configuring-search) for the full reference. ```python theme={null} from perplexity import Perplexity client = Perplexity() # Override the model while keeping everything else from the preset response = client.responses.create( preset="pro-search", model="anthropic/claude-sonnet-4-6", # Use Claude instead of the default GPT-5.1 input="Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", ) # Override max_steps for deeper reasoning response = client.responses.create( preset="pro-search", input="What is serverless cold start latency, what causes it, and what are the standard mitigations (warm pools, provisioned concurrency)?", max_steps=5, # Override preset's default of 3 ) # Override tools configuration with a static search config response = client.responses.create( preset="pro-search", input="Explain the FDA's accelerated approval pathway under 21 CFR 314 Subpart H: eligibility criteria, surrogate endpoints, and confirmatory trial requirements.", tools=[{ "type": "web_search", "search_context_size": "high", "filters": { "search_domain_filter": ["clinicaltrials.gov", "fda.gov"], # Restrict to specific domains }, }], ) # Use explicit token budgets when you need exact budget control response = client.responses.create( preset="pro-search", input="Explain the FDA's accelerated approval pathway under 21 CFR 314 Subpart H: eligibility criteria, surrogate endpoints, and confirmatory trial requirements.", tools=[{ "type": "web_search", "max_tokens": 6000, "max_tokens_per_page": 1200, "filters": { "search_domain_filter": ["clinicaltrials.gov", "fda.gov"], }, }], ) ``` ```typescript theme={null} import Perplexity from '@perplexity-ai/perplexity_ai'; const client = new Perplexity(); // Override the model while keeping everything else from the preset const response = await client.responses.create({ preset: "pro-search", model: "anthropic/claude-sonnet-4-6", // Use Claude instead of the default GPT-5.1 input: "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP.", }); // Override max_steps for deeper reasoning const response2 = await client.responses.create({ preset: "pro-search", input: "What is serverless cold start latency, what causes it, and what are the standard mitigations (warm pools, provisioned concurrency)?", max_steps: 5, // Override preset's default of 3 }); // Override tools configuration with a static search config const response3 = await client.responses.create({ preset: "pro-search", input: "Explain the FDA's accelerated approval pathway under 21 CFR 314 Subpart H: eligibility criteria, surrogate endpoints, and confirmatory trial requirements.", tools: [{ type: "web_search" as const, search_context_size: "high", filters: { search_domain_filter: ["clinicaltrials.gov", "fda.gov"], // Restrict to specific domains }, }], }); // Use explicit token budgets when you need exact budget control const response4 = await client.responses.create({ preset: "pro-search", input: "Explain the FDA's accelerated approval pathway under 21 CFR 314 Subpart H: eligibility criteria, surrogate endpoints, and confirmatory trial requirements.", tools: [{ type: "web_search" as const, max_tokens: 6000, max_tokens_per_page: 1200, filters: { search_domain_filter: ["clinicaltrials.gov", "fda.gov"], }, }], }); ``` ```bash theme={null} # Override the model while keeping everything else from the preset curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "pro-search", "model": "anthropic/claude-sonnet-4-6", "input": "Summarize the core findings of the original 'Attention Is All You Need' transformer paper and explain why it changed NLP." }' | jq ``` ```bash theme={null} # Override max_steps for deeper reasoning curl https://api.perplexity.ai/v1/agent \ -H "Authorization: Bearer $PERPLEXITY_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "preset": "pro-search", "input": "What is serverless cold start latency, what causes it, and what are the standard mitigations (warm pools, provisioned concurrency)?", "max_steps": 5 }' | jq ``` ```json theme={null} { "id": "resp_e80c520d-2788-4b3d-8662-6a0e3d611e8f", "created_at": 1779391438, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "Abstract:The dominant sequence transduction models are based on complex recurrent or convolutional neural networks in an encoder-decoder configuration.\nThe best performing models also connect the encoder and decoder through an attention mechanism.\nWe propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely.\nExperiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.\nOur model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles by over 2 BLEU.\nOn the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature.\nWe show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.", "title": "[1706.03762] Attention Is All You Need - arXiv", "url": "https://arxiv.org/abs/1706.03762", "date": "2017-06-12", "last_updated": "2026-05-19", "source": "web" }, { "id": 2, "snippet": "At each layer, each token is then contextualized within the scope of the context window with other (unmasked) tokens via a parallel multi-head attention mechanism, allowing the signal for key tokens to be amplified and less important tokens to be diminished.\n...\nThe original version of the transformer architecture was proposed in the 2017 paper \"Attention Is All You Need\" by researchers at Google.\nThe predecessors of transformers were developed as an improvement over previous architectures for machine translation, but have found many applications since.\n...\nOn 2017-06-12, the original (100M-parameter) encoder–decoder transformer model was published in the \"Attention is all you need\" paper.\nAt the time, the focus of the research was on improving seq2seq for machine translation, by removing its recurrence to process all tokens in parallel, but preserving its dot-product attention mechanism to keep its text processing performance.\nThis led to the introduction of a multi-head attention model that was easier to parallelize due to the use of independent heads and the lack of recurrence.\nIts parallelizability was an important factor to its widespread use in large neural networks.", "title": "Transformer (deep learning) - Wikipedia", "url": "https://en.wikipedia.org/wiki/Transformer_(deep_learning)", "date": "2019-08-25", "last_updated": "2026-05-15", "source": "web" }, { "id": 3, "snippet": "We propose a novel, simple network architecture based solely onan attention mechanism, dispensing with recurrence and convolutions entirely.Experiments on two machine translation tasks show these models to be superiorin quality while being more parallelizable and requiring significantly less timeto train.", "title": "Attention is All you Need - NIPS papers", "url": "https://papers.nips.cc/paper/7181-attention-is-all-you-need", "date": "2017-01-01", "last_updated": "2026-05-21", "source": "web" }, { "id": 4, "snippet": "The paper introduced a new deep learning architecture known as the transformer, based on the attention mechanism proposed in 2014 by Bahdanau *et al.* The transformer approach it describes has become the main architecture of a wide variety of artificial intelligence, including large language models.\nAt the time, the focus of the research was on improving Seq2seq techniques for machine translation, but the authors go further in the paper, foreseeing the technique's potential for other tasks like question answering and what is now known as multimodal generative AI.\n...\nThe paper is best known for introducing the Transformer architecture, which underlies most modern large language models (LLMs).\nA key reason why the architecture is preferred by most modern LLMs is the parallelizability of the architecture over its predecessors.\nThis ensures that the operations necessary for training can be accelerated on a GPU, allowing both faster training times and models of bigger sizes to be trained.\n...\nOn 2017-06-12, the original (100M-parameter) encoder–decoder transformer model was published in the \"Attention is all you need\" paper.", "title": "Attention Is All You Need - Wikipedia", "url": "https://en.wikipedia.org/wiki/Attention_Is_All_You_Need", "date": "2023-12-04", "last_updated": "2026-05-17", "source": "web" }, { "id": 5, "snippet": "In this video, YC's Ankit Gupta traces how AI learned to understand language — from early RNNs and LSTMs to attention mechanisms and the breakthrough 2017 paper Attention Is All You Need — the discovery that unlocked the modern AI era.\n...\n{ts:17} [Music] A transformer is a neural network that uses self attention to take input data\n{ts:26} like text or images, model the relationships between that data, and finally generate outputs like meaningful\n{ts:32} text responses, translations, or classifications.\n...\nThen came the big breakthrough in 2017 when a team of researchers at\n{ts:387} Google published a paper called attention is all you need, which proposed a new machine translation\n{ts:392} architecture that they called a transformer.\n...\n{ts:409} version of the encoder decoder architecture originally proposed in seek to seek.\nInstead of compressing inputs\n{ts:414} into a single vector embedding, transformers kept separate embeddings for each input token and updated these\n{ts:420} through self attention, a mechanism that updated token representations based on a learned weighted dotproduct over the\n{ts:426} embeddings of all other tokens in the sequence.\nBecause each token in this architecture could attend to all others\n{ts:431} simultaneously, transformers could process an entire sequence in parallel, making them dramatically faster than\n{ts:437} RNN's.\nRemarkably, they were also much more accurate on machine translation benchmarks.", "title": "Transformers Explained: The Discovery That Changed AI Forever", "url": "https://www.youtube.com/watch?v=JZLZQVmfGn8&vl=en-US", "date": "2025-10-23", "last_updated": "2026-03-27", "source": "web" }, { "id": 6, "snippet": "●Presents a new neural architecture named the Transformer\n●Based solely on the attention mechanism widely used in SEQ2SEQ models\n●More parallelizable compared to existing state-of-the-art (SOTA) models\n●Achieves SOTA in 2 machine translation datasets\n…\n...\n• Less total computational complexity per layer\n• More parallelizable than existing fully autoregressive models\n• Shorten the path between tokens to enable model to learn long-term \ndependency better\n...\nIntroduces a groundbreaking new model that is solely based on attention\n●\nFaster and better than existing models\n●", "title": "[PDF] Attention Is All You Need", "url": "https://ysu1989.github.io/courses/au20/cse5539/Transformer.pdf", "date": null, "last_updated": "2026-05-15", "source": "web" }, { "id": 7, "snippet": "The Transformer architecture has revolutionized natural language processing (NLP) since its introduction, establishing itself as a cornerstone for modern advancements in the field.\nThis architecture offers significant improvements in handling a wide range of NLP tasks, from translation to text summarization, by overcoming limitations inherent in previous models.\nIntroduced by Vaswani et al. in the paper “Attention is All You Need” in 2017, the Transformer architecture is characterized by its use of self-attention mechanisms and feed-forward neural networks.\nUnlike its predecessors, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), which process data sequentially, the Transformer processes input data in parallel.\nThis parallelization enables the model to leverage modern hardware more efficiently, resulting in faster training times and improved scalability.\nAt the heart of the Transformer is the self-attention mechanism, which allows the model to weigh the significance of different words in a sentence relative to each other.\nThis feature is critical for understanding contextual relationships and dependencies, regardless of the distance between words in the input sequence.\n...\nThis capacity to capture global dependencies without the constraints of sequential data processing is a notable advantage over traditional models.\n...\nThe Transformer architecture’s flexibility makes it suitable for a broad spectrum of NLP applications.\nIn machine translation, it has significantly improved the accuracy and fluency of translated text.\nFor tasks like text summarization, sentiment analysis, and question answering, Transformers provide state-of-the-art results by effectively capturing and modeling complex language patterns.\n...\nIn summary, the Transformer architecture has fundamentally changed the landscape of NLP by introducing a novel approach to processing and understanding language data.\nIts innovative use of self-attention and parallel processing has paved the way for numerous breakthroughs, making it an essential tool for anyone looking to advance in the field of natural language processing.", "title": "What is the Transformer architecture in NLP? - Milvus", "url": "https://milvus.io/ai-quick-reference/what-is-the-transformer-architecture-in-nlp", "date": "2026-03-26", "last_updated": "2026-05-13", "source": "web" }, { "id": 8, "snippet": "In “Attention Is All You Need”, we introduce the Transformer, a novel neural network architecture based on a self-attention mechanism that we believe to be particularly well suited for language understanding.\nIn our paper, we show that the Transformer outperforms both recurrent and convolutional models on academic English to German and English to French translation benchmarks.\nOn top of higher translation quality, the Transformer requires less computation to train and is a much better fit for modern machine learning hardware, speeding up training by up to an order of magnitude.\n...\nIn contrast, the Transformer only performs a small, constant number of steps (chosen empirically).\nIn each step, it applies a self-attention mechanism which directly models relationships between all words in a sentence, regardless of their respective position.", "title": "Transformer: A Novel Neural Network Architecture for Language ...", "url": "https://research.google/blog/transformer-a-novel-neural-network-architecture-for-language-understanding/", "date": null, "last_updated": "2026-05-17", "source": "web" }, { "id": 9, "snippet": "A new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely is proposed, which generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.", "title": "[PDF] Attention is All you Need | Semantic Scholar", "url": "https://www.semanticscholar.org/paper/Attention-is-All-you-Need-Vaswani-Shazeer/204e3073870fae3d05bcbc2f6a8e263d9b72e776", "date": "2017-06-12", "last_updated": "2025-03-22", "source": "web" }, { "id": 10, "snippet": "Abstract:Natural Language Processing (NLP) has witnessed a transformative leap with the advent of transformer-based architectures, which have significantly enhanced the ability of machines to understand and generate human-like text.\nThis paper explores the advancements in transformer models, such as BERT and GPT, focusing on their superior performance in text understanding tasks compared to traditional methods like recurrent neural networks (RNNs).\nBy analyzing statistical properties through visual representations-including probability density functions of text length distributions and feature space classifications-the study highlights the models' proficiency in handling long-range dependencies, adapting to conditional shifts, and extracting features for classification, even with overlapping classes.\nDrawing on recent 2024 research, including enhancements in multi-hop knowledge graph reasoning and context-aware chat interactions, the paper outlines a methodology involving data preparation, model selection, pretraining, fine-tuning, and evaluation.\nThe results demonstrate state-of-the-art performance on benchmarks like GLUE and SQuAD, with F1 scores exceeding 90%, though challenges such as high computational costs persist.\nThis work underscores the pivotal role of transformers in modern NLP and suggests future directions, including efficiency optimization and multimodal integration, to further advance language-based AI systems.", "title": "Exploring Transformer-Based Architectures for Text Understanding", "url": "https://arxiv.org/abs/2503.20227", "date": "2025-03-26", "last_updated": "2026-02-21", "source": "web" } ], "type": "search_results", "queries": [ "Attention Is All You Need transformer paper core contributions", "why transformer architecture changed NLP", "Vaswani et al 2017 attention is all you need key results" ] }, { "id": "msg_b9dc751b-50fd-4941-ad3d-00bb2eb98f73", "content": [ { "text": "The paper “Attention Is All You Need” introduced the **Transformer**, a sequence‑to‑sequence model built entirely from attention and feed‑forward layers, and showed it could beat recurrent and convolutional models on machine translation while training much faster and scaling better.[web:1][web:8] This architectural shift made self‑attention the default building block for modern NLP and enabled today’s large language models.[web:2][web:4][web:7]\n\n## Core technical ideas\n\n- The Transformer removes recurrence and convolutions, using stacked self‑attention and position‑wise feed‑forward layers in both encoder and decoder.[web:1][web:3] Positional encodings are added to token embeddings so the model can represent order despite having no recurrence.[web:1] \n- It introduces **multi‑head** attention: several parallel attention heads project queries/keys/values, attend in different subspaces, then concatenate and linearly mix their outputs.[web:1][web:2] This lets the model capture different types of relations (e.g., syntax and coreference) in parallel.[web:1] \n- The paper emphasizes three attention types within the architecture: encoder self‑attention, decoder masked self‑attention (preventing access to future tokens), and encoder‑decoder attention where the decoder attends over encoder outputs.[web:1][web:8]\n\n## Empirical results in the paper\n\n- On WMT 2014 English→German translation, the base Transformer reaches 28.4 BLEU, surpassing prior state of the art, including ensembles, by over 2 BLEU.[web:1][web:3] On WMT 2014 English→French, a single Transformer achieves 41.8 BLEU after 3.5 days of training on 8 GPUs, outperforming earlier, much more expensive systems.[web:1] \n- The authors also show the model generalizes beyond translation by achieving strong results on English constituency parsing, even with limited labeled data.[web:1][web:9] These results demonstrate that attention‑only architectures are not just viable but competitive across sequence tasks.[web:1]\n\n## Why it was faster and more scalable\n\n- Self‑attention lets the model process all tokens in a sequence in parallel, unlike RNNs/LSTMs that must step through tokens sequentially, which makes it far more compatible with GPU/TPU hardware.[web:1][web:7] This parallelism reduced training time “by up to an order of magnitude” relative to recurrent models on the same tasks.[web:8] \n- For typical sequence lengths, the attention‑based layers in the paper have lower or comparable per‑layer computational complexity than recurrent or convolutional alternatives and much shorter paths between distant tokens, which helps learn long‑range dependencies.[web:1][web:6] This combination of efficiency and representational power is what made large‑scale pretraining feasible.[web:7][web:10]\n\n## How it changed NLP\n\n- The Transformer architecture became the foundation for major pretrained models such as BERT (encoder‑only), GPT (decoder‑only), and many others, which now dominate benchmarks in translation, question answering, summarization, and general language modeling.[web:2][web:7][web:10] These models exploit the same self‑attention blocks introduced in the 2017 paper, often with minor variants.[web:2] \n- By showing that attention‑only, highly parallel architectures can outperform traditional sequence models, the paper shifted the field’s default design pattern from RNNs/CNNs to transformers and opened the path to scaling parameters and data to today’s very large language models.[web:2][web:4][web:7] Its influence now extends beyond NLP to vision, speech, and multimodal models that reuse the same core transformer building blocks.[web:2][web:4]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 6620, "output_tokens": 863, "total_tokens": 7483, "cost": { "currency": "USD", "input_cost": 0.0038, "output_cost": 0.00863, "total_cost": 0.01788, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391438, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` ````json theme={null} { "id": "resp_cd7d548e-e9a9-4eeb-8dd0-e68bf15848bb", "created_at": 1779391825, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "## What is a cold start in serverless computing?compIn the context of serverless computing, a “cold start” refers to the first-time latency or the time it takes for a function to begin processing after it has not been called in a long time.\nTo illustrate, when a computer is turned on after being shut down, it may need some time to reach its full performance.", "title": "What is Cold Start? Understanding Serverless Latency", "url": "https://payproglobal.com/answers/what-is-cold-start/", "date": "2026-04-02", "last_updated": "2026-05-19", "source": "web" }, { "id": 2, "snippet": "A Lambda cold start occurs when a new function instance must be created and initialized.\nThe cold start refers to the delay between invocation and runtime created by the initialization process.\nLambda cold starts occur when there is no available function instance to respond to an invocation.\nThis can happen when instances have expired due to inactivity or when there are more invocations than active instances.\nCold starts are an inherent problem with serverless models because providers are unwilling or unable to keep inactive instances alive indefinitely.\n...\nLambda can take up to 10 seconds to complete this process.\n...\nA Lambda cold start occurs when function instances are not available to respond to invocations.\nThere are several factors contributing to this issue.\n...\nIf you are planning on setting up your Lambda on VPC, you need to be aware that the process of initializing Lambda functions on VPCs can take up to 10 seconds.", "title": "5 Ways to Manage Lambda Cold Starts | Khalil Stemmler", "url": "https://khalilstemmler.com/blogs/serverless/5-ways-to-manage-lambda-cold-starts/", "date": "2020-06-27", "last_updated": "2026-03-21", "source": "web" }, { "id": 3, "snippet": "Provisioned concurrency – This is the number of pre-initialized execution\nenvironments allocated to your function.\nThese execution environments are ready to respond\nimmediately to incoming function requests.\nProvisioned concurrency is useful for reducing\ncold start latencies for functions and designed to make functions available with double-digit\nmillisecond response times.", "title": "Configuring provisioned concurrency for a function - AWS Lambda", "url": "https://docs.aws.amazon.com/lambda/latest/dg/provisioned-concurrency.html", "date": "2023-01-21", "last_updated": "2026-05-02", "source": "web" }, { "id": 4, "snippet": "Usually, a cold start is a term used to define a situation where applications take longer than expected to start up and respond to a request.\nIn the serverless context, a cold start is the total time that users must wait to get a response for their function request.\n...\nIf there are no available containers, it has to create a new one, which is referred to as a cold start.\nEffectively, on invoking a serverless function (in a cold state), every request takes additional time to be processed, thus resulting in high latency.", "title": "Serverless Cold Starts–Latency Reducing Strategies | Wissen", "url": "https://www.wissen.com/blog/serverless-cold-starts-latency-reducing-strategies", "date": "2025-02-03", "last_updated": "2026-05-21", "source": "web" }, { "id": 5, "snippet": "AWS needs a ready supply of containers to spin up when functions are invoked.\nThat means that functions are kept warm for a limited amount of time (usually 30 – 45 minutes) after executing, before being spun down so that container is ready for any new function to be invoked.\n...\nCold starts in Lambda account for less than **0.25%** of requests but the impact can be huge, sometimes requiring 5 seconds to execute the code.\nThis issue is particularly relevant to applications that need to run executions in real-time, or those that rely on split-second timing.\n...\nIf one function in the chain experiences a cold start, it can delay the entire chain of functions, leading to increased overall latency.", "title": "What Causes AWS Lambda Cold Starts & 7 Ways to Solve Them", "url": "https://lumigo.io/blog/this-is-all-you-need-to-know-about-lambda-cold-starts/", "date": "2020-09-20", "last_updated": "2026-05-12", "source": "web" }, { "id": 6, "snippet": "A cold start occurs when a Lambda function is invoked for the first time or after a period of inactivity.\nAWS Lambda must create a new execution environment, which involves:\n- Downloading the function’s code from an internal S3 bucket.\n- Setting up the runtime environment, including operating system and language dependencies.\n- Initializing the function’s code and any external libraries.\nThis process introduces latency, with durations typically ranging from under 100 milliseconds to over 1 second, depending on factors like runtime, package size, and memory allocation.\nAccording to AWS documentation, cold starts occur in less than 1% of invocations, but their impact can be significant in latency-sensitive applications AWS Lambda Documentation.\n...\nCold starts introduce latency that can degrade user experience, particularly in applications requiring real-time responses, such as APIs, authentication services, or e-commerce platforms.\n...\nCold starts in AWS Lambda can introduce significant latency, impacting user experience in latency-sensitive applications.", "title": "AWS Lambda Cold and Warm Starts: Impacts and Mitigation Strategies", "url": "https://dev.to/mystyx/aws-lambda-cold-and-warm-starts-impacts-and-mitigation-strategies-354j", "date": "2025-07-05", "last_updated": "2025-09-19", "source": "web" }, { "id": 7, "snippet": "Cold starts are an important consideration when building applications on serverless platforms.\nIn AWS Lambda, they refer to the initialization steps that occur when a function is invoked after a period of inactivity or during rapid scale-up.\nWhile typically brief and infrequent, cold starts can introduce additional latency, making it essential to understand them, especially when optimizing performance in responsive and latency-sensitive workloads.\n...\nCold starts occur because serverless platforms like AWS Lambda are designed for cost-efficiency – you don’t pay for compute resources when your code isn’t running.\nAs a result, Lambda only provisions resources when needed.\nA cold start happens when there isn’t an existing execution environment available and a new one must be created.\nThis can happen, for example, when a function is invoked for the first time after a period of inactivity or during a burst in traffic that triggers scale-up.\nWhen this occurs, Lambda rapidly provisions and initializes a new execution environment for running your function code.\nThis initialization adds a small amount of latency to the request, but it only occurs once for the lifecycle of that execution environment.\nCold starts consist of several steps that make up the **Initialization Phase**, which occurs before the function begins running.\nThese steps take place when the Lambda service creates a new execution environment, contributing to the latency commonly referred to as the INIT duration of the function, as illustrated in a following diagram:\n...\nOptimizing your function initialization code can help to reduce cold start latencies.\nStreamline your function code to load and prepare quickly, alongside its runtime environment and dependencies.\nEmploy lightweight libraries and implement lazy loading for resources to further cut initialization time.\nMinimize code size by eliminating unnecessary dependencies.\nConsider your architecture carefully: break down large functions into smaller, more focused units based on invocation patterns.", "title": "Understanding and Remediating Cold Starts: An AWS Lambda ...", "url": "https://aws.amazon.com/blogs/compute/understanding-and-remediating-cold-starts-an-aws-lambda-perspective/", "date": "2025-08-07", "last_updated": "2026-05-21", "source": "web" }, { "id": 8, "snippet": "When you invoke a serverless function that hasn’t run in a while, the cloud provider has to go through several steps:\n```\n1. Download Function Code (from S3 or container registry)\n2. Initialize Execution Environment (spin up container/VM)\n3. Load Runtime (Node.js, Python, Java, etc.)\n4. Import Dependencies (your libraries and modules)\n5. Execute Initialization Code (global variables, DB connections)\n6. Finally: Run Your Handler Function\n```\nSteps 1–5 are what we call the “cold start.”\nYour code only begins executing at step 6.\nFor a simple “Hello World” function, this might take 200–500ms.\nBut for real-world applications with dependencies?\nI’ve seen cold starts take 5+ seconds, especially with Java or .NET runtimes.\n...\nCold start times of a WebAssembly runtime are under one millisecond, making it orders of magnitude faster than Docker containers, virtual machines, or even modern cloud-based serverless functions.\n...\nThe key is choosing the right strategy for each function based on its usage patterns, latency requirements, and your budget constraints.", "title": "Why Serverless Cold Starts Happen (And the 3 Ways to ...", "url": "https://medium.com/@sohail_saifi/why-serverless-cold-starts-happen-and-the-3-ways-to-eliminate-them-a3ddf3890067", "date": "2025-07-05", "last_updated": "2025-10-11", "source": "web" }, { "id": 9, "snippet": "So this is something that can be useful for reducing cold start latencies for a function.\n...\nLuciano: Yeah, provisioned concurrency, as you said, keeps a certain number of Lambda execution environments warm for you.\nSo this basically means that as soon as you have enabled provisioned concurrency and set a specific amount for a function, AWS will need to spin up that number of execution environments for you so that they are ready and warm for whenever new events come in.\nSo basically if you receive a request, you will have this Lambda environment already available.\nAnd also this environment not going to be eventually disposed by AWS, even though you might have a period of time where you don't receive enough events, or maybe you have even zero traffic.\nIf you have provisioned concurrency, your instances will still be there and available, even if nothing happening in your account.\nSo in a way, this is going to help you to fight cold starts, but it doesn't necessarily mean that you won't have cold starts anymore.\n...\nAnd that means that those new instances will incur in a cold start.", "title": "Lambda Provisioned Concurrency - AWS Bites podcast", "url": "https://awsbites.com/129-lambda-provisioned-concurrency/", "date": "2024-08-23", "last_updated": "2026-05-19", "source": "web" }, { "id": 10, "snippet": "This process is known as scaling to zero [Castro et al., 2019].\nResources scaled to zero take some time to be processed for reuse when a new function arrives.\nThis latency is called a cold start in serverless computing [Baldini et al., 2017].\nCold start latency has negative effects such as delayed response to time-sensitive applications, inconsistent performance, increased resource consumption, and poor user experience [Wen et al., 2023].\n...\n*Latency*: It is defined as the time delay between consumer request submission and receiving the response from the corresponding serverless function [Golec et al., 2023b, Sedlak et al., 2024].\nThis time is resultant of request queuing time, request-to-resource mapping time, and finally the serverless function execution time.\nHence, reducing the cold start delay in resource set-up significantly optimizes this parameter making it suitable for real-life use cases.\n...\nAfter the containers are scaled to zero, it is necessary to start the container again for the requests coming to the server.\nStarting a new container and preparing the function for execution causes a specific latency.\nIn the serverless paradigm, this latency period is called a cold start [Kim and Lee, 2019].", "title": "Cold Start Latency in Serverless Computing: A Systematic Review ...", "url": "https://arxiv.org/html/2310.08437v2", "date": null, "last_updated": "2026-02-18", "source": "web" }, { "id": 11, "snippet": "Cold starts occur when a serverless function needs to initialize a new execution environment before processing a request.\nThis adds latency that can range from a few hundred milliseconds to several seconds.", "title": "How to Fix 'Cold Start' Serverless Issues - OneUptime", "url": "https://oneuptime.com/blog/post/2026-01-24-fix-cold-start-serverless-issues/view", "date": "2026-01-24", "last_updated": "2026-05-17", "source": "web" }, { "id": 12, "snippet": "AWS Lambda cold starts (the time it takes for AWS to assign a worker to a request) are a major frustration point of many serverless programmers.\n...\nWhenever Lambda receives a request but it has no idle workers, the control plane assigns a new generic worker to it.\nThe worker then has to download the custom code or binaries of your Lambda and load them into memory before it can service the request.\nThis process takes time, which significantly increases response latency.\nThe issue of sporadically slow responses caused by the need to increase the pool of workers is known as **Cold Start**.\nCold starts are consistently the top concern about the applicability of serverless technologies to latency-sensitive workloads.", "title": "Provisioned Concurrency: Avoiding Cold Starts in AWS Lambda", "url": "https://www.pulumi.com/blog/aws-lambda-provisioned-concurrency-no-cold-starts/", "date": "2019-12-19", "last_updated": "2026-05-16", "source": "web" }, { "id": 13, "snippet": "For each function, users specify triggers, such as timers,\nthat activate the function.\nWhen a function is triggered, a\ncontainer with that function’s code processes the request.\nIf a new container needs to be started either because there\nare no active containers for this function or because existing\ncontainers are overloaded, this causes a cold start.\nThis cold\nstart adds significant latency, degrading application perfor-\nmance.\n...\nOne of the main bottlenecks of serverless systems is cold\nstarts, where the system must start a new pod if it does not\nhave sufficient resources to process an incoming request [18,\n23, 39].", "title": "[PDF] Serverless Cold Starts and Where to Find Them", "url": "http://faculty.washington.edu/wlloyd/courses/tcss562/papers/2024/ServerlessColdStartsandWheretoFindThem.pdf", "date": null, "last_updated": "2026-03-22", "source": "web" }, { "id": 14, "snippet": "Instances of cloud functions are added and removed dynamically.\nWhen a new instance handles its first request, the response time increases, which is called a **cold start**.\n...\nAWS clearly leads with all languages being **well below 1 second**.\nGCP start-up usually takes **between 0.5 and 2 seconds**.\nAzure is a clear underdog with startup times often **up to 5 seconds**.", "title": "Comparison of Cold Starts in Serverless Functions across AWS ...", "url": "https://mikhail.io/serverless/coldstarts/big3/", "date": "2021-01-05", "last_updated": "2026-02-22", "source": "web" } ], "type": "search_results", "queries": [ "serverless cold start latency definition", "causes of cold starts serverless functions", "provisioned concurrency warm pools mitigation cold start AWS Lambda Azure Functions Cloud Functions" ] }, { "id": "msg_e93531f0-d063-4588-9959-9e634ae7f17a", "content": [ { "text": "Serverless cold start latency is the extra delay a request experiences while the platform spins up a new execution environment for a function that has no warm instance available. [web:10][web:13] It is caused by “scale-to-zero” and rapid scale‑out behavior plus function initialization work, and is usually mitigated with warm pools and configuration features like provisioned concurrency that keep some instances pre‑initialized. [web:7][web:3][web:12]\n\n## What cold start latency is\n\n- In serverless platforms, a cold start is the time added to a request when a new container/VM must be started to run a function because no suitable instance is currently running. [web:10][web:13] \n- Measured at the user level, cold start latency is the extra time between sending a request and receiving the first response compared to a warm invocation of the same function. [web:4][web:11] \n\nTypical cold starts can range from tens of milliseconds up to several seconds, depending on runtime, package size, and platform. [web:6][web:14]\n\n## What actually happens during a cold start\n\nWhen a function is invoked “cold,” the platform must go through a series of initialization steps before your handler code runs. [web:7][web:8]\n\nCommon steps include:\n\n- Allocating or starting an execution environment (container or lightweight VM) for the function. [web:7][web:10] \n- Downloading or mounting the function package or image from internal storage or a registry. [web:6][web:12] \n- Initializing the runtime (Node.js, Python, Java, .NET, etc.) and loading your code and its dependencies. [web:6][web:8] \n- Running global or initialization code, such as establishing DB connections or loading configuration. [web:6][web:7] \n\nOnly after these steps does the platform actually invoke the function handler with the event, which is why the first request is slower than subsequent ones in the same environment. [web:6][web:8]\n\n## Main causes of cold starts\n\nCold starts are primarily a side effect of the serverless resource model, where providers scale instances up and down aggressively to save cost and isolate tenants. [web:7][web:10]\n\nKey causes:\n\n- **Scale‑to‑zero and idle reaping**: After a function is idle for some period, its containers are stopped; the next invocation needs a fresh environment. [web:5][web:10] \n- Burst scaling: Traffic spikes can require more instances than are currently warm, forcing the platform to create new ones. [web:7][web:13] \n- Heavy initialization: Large deployment packages, many libraries, synchronous imports, and complex startup logic increase initialization time. [web:6][web:7] \n- Runtime and networking overhead: Some runtimes (e.g., Java, .NET) and configurations (e.g., VPC-attached Lambdas) have extra startup and network setup cost. [web:2][web:6] \n\nThese factors make cold starts especially visible in latency‑sensitive APIs and user‑facing paths. [web:6][web:5]\n\n## Warm pools and keeping functions warm\n\nA warm pool is a set of pre‑created execution environments that are kept alive so they can handle requests without paying full cold start cost. [web:12][web:9]\n\nThe concept shows up in a few forms:\n\n- Provider‑managed idle instances: Platforms maintain some number of idle containers per function based on recent traffic; if a request hits an idle instance, it’s effectively warm. [web:7][web:13] \n- Explicit “keep warm” patterns: Users schedule periodic invocations (e.g., cron triggers) so that environments don’t get torn down, reducing the number of true cold invocations. [web:2][web:6] \n- Warm pools for scale‑out: Some systems maintain a generic pool of pre‑booted workers that only need the function code attached, reducing but not eliminating cold start latency. [web:12][web:10] \n\nWarm pools mitigate latency by reusing initialized environments, but they cannot fully prevent cold starts during sharp scale‑out or if traffic patterns differ from the warm‑up strategy. [web:9][web:13]\n\n## Provisioned concurrency and similar features\n\nProvisioned concurrency (AWS Lambda) and analogous features on other platforms formalize warm pools by letting you reserve a fixed number of pre‑initialized instances. [web:3][web:9]\n\nCore properties:\n\n- A configured number of execution environments is created ahead of time, with runtime and your code already initialized and ready to handle events. [web:3][web:9] \n- Requests routed to these provisioned environments have consistently low, “double‑digit millisecond” level startup overhead, effectively eliminating cold start latency for that portion of traffic. [web:3][web:12] \n\nImportant nuances:\n\n- If concurrent traffic exceeds the provisioned level, extra instances will still be created and those invocations will incur cold starts. [web:9][web:7] \n- You pay for the reserved capacity while it’s provisioned, whether or not it is actively handling requests, trading off cost for predictable latency. [web:3][web:12] \n\nIn practice, platforms combine provisioned concurrency or similar reserved‑capacity features with autoscaling and implicit warm pools to balance cost efficiency with reduced cold start latency. [web:7][web:10]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 7278, "output_tokens": 1214, "total_tokens": 8492, "cost": { "currency": "USD", "input_cost": 0.00462, "output_cost": 0.01214, "total_cost": 0.02221, "cache_creation_cost": null, "cache_read_cost": 0.00045, "tool_calls_cost": 0.005 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 3584, "cached_tokens": 3584 }, "tool_calls_details": { "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391825, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ```` ```json theme={null} { "id": "resp_c06bc7e4-0aed-4650-b5f1-712f12067af1", "created_at": 1779391734, "model": "openai/gpt-5.1", "object": "response", "output": [ { "results": [ { "id": 1, "snippet": "2. § 314.510 Approval based on a surrogate endpoint or on an effect on a clinical endpoint other than survival or irreversible morbidity.", "title": "21 CFR Part 314 - Subpart H - Accelerated Approval of New Drugs ...", "url": "https://www.law.cornell.edu/cfr/text/21/part-314/subpart-H", "date": null, "last_updated": "2026-05-16", "source": "web" }, { "id": 2, "snippet": "^^ A validated surrogate endpoint “is known to predict clinical benefit” for a certain disease state and for a certain type of intervention.^^ It has been suggested that to be a validated surrogate endpoint, the biological marker “must be correlated with the clinical endpoint” and “must fully capture the net effect of the intervention on the clinical-efficacy endpoint” for a specific disease setting and class of interventions.\n...\nFor a new compound to qualify for Accelerated Approval^^ following the enactment of FDASIA, it must address a serious or life-threatening condition and demonstrate an effect on a surrogate endpoint or an intermediate clinical endpoint other than a direct measure of mortality or survival.^^ The surrogate endpoint or intermediate clinical endpoint must be “reasonably likely to predict” the clinical benefit or an effect on irreversible morbidity or mortality.^^\n...\n…\nFDA defines a “surrogate endpoint” in its Final Guidance as “a marker, such as a laboratory measurement, radiographic image, physical sign, or other measure, that is thought to predict clinical benefit, but is not itself a measure of clinical benefit.”", "title": "FDA's Expedited Approval Mechanisms for New Drug Products - PMC", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4326266/", "date": "2015-02-01", "last_updated": "2026-03-25", "source": "web" }, { "id": 3, "snippet": "## BackgroundgrAccelerated approval is intended to facilitate quicker development of drugs intended to treat serious and life-threatening conditions in areas of unmet medical need.\nSpecifically, by using accelerated approval, FDA can meet the approval standard for drugs and biological products based on data demonstrating an impact on a surrogate endpoint or an intermediate clinical endpoint that is reasonably likely to predict clinical benefit.\nBy relying on endpoints that can be met earlier than clinical endpoints, sponsors and FDA can speed treatments to patients who need them and push additional work — that of verifying that the surrogate or intermediate endpoint is in fact predictive of clinical benefit — to post-marketing confirmatory studies.\n...\nA surrogate endpoint that would be appropriate for accelerated approval, the Draft Guidance explains, can be a biomarker that is reasonably likely to predict a clinical benefit.\nA surrogate endpoint that is known to predict clinical benefit, however, would support a traditional (not accelerated) approval.\nAnd, in the absence of sufficient evidence to support reliance on a biomarker as having either a known or predicted impact on clinical benefit, neither traditional nor accelerated approval would be supportable.\n...\nAccelerated approval can also be supported by an intermediate clinical endpoint, i.e., a measurement of a therapeutic effect that can be measured earlier than irreversible morbidity or mortality — except that sponsors should first consider “whether the demonstrated therapeutic effect on the intermediate clinical endpoint alone would be a basis for traditional approval,” in which case the traditional approval pathway should be used.\nApproval based on intermediate clinical endpoints will be considered for accelerated approval only if it is critical to confirm the effects on irreversible morbidity or mortality (IMM) or other clinical benefits.\n...\nIn addition to sufficient evidence of safety and effectiveness, the Draft Guidance explains, an application for accelerated approval should include adequate evidence that the proposed surrogate or intermediate clinical endpoint is reasonably likely to predict clinical benefit.\nWhether or not a surrogate or intermediate clinical endpoint is reasonably likely to predict clinical benefit often requires a judgment call in conjunction with empirical evidence to support that relationship.\nWith the caveat that the specific clinical evidence needed to support such a determination will be case-specific, the Draft Guidance outlines several important factors that will inform that judgment, including:\n- How well the drug’s effectiveness on a clinical endpoint is understood.\nThe clearer and more well-known the relationship, the likelier it is that the drug should go through the traditional approval process.\n- Whether there is reliable and consistent evidence to support the correlation between endpoint and clinical outcome of interest with a particular attention placed on the source and nature of the evidence.\n- Whether there is evidence from a clinical trial supporting that the effect on the surrogate endpoint has been shown to predict a clinical benefit with another drug.\nFDA emphasizes its expectation that “[c]linical data should be provided” to support the relationship.^*11* ^\n...\nAlso of note is the focus in the Draft Guidance on the idea that validated surrogates should be used to seek traditional, not accelerated, approval.^*16* ^\n…\n...\n2. 21 C.F.R.\n§§ 314.500 et seq.", "title": "FDA Issues New Draft Guidance for Expedited Program for Serious ...", "url": "https://www.arnoldporter.com/en/perspectives/advisories/2024/12/fda-draft-guidance-expedited-program-for-serious-conditions", "date": "2024-12-12", "last_updated": "2026-03-24", "source": "web" }, { "id": 4, "snippet": "The FDA instituted its Accelerated Approval Program to allow for earlier approval of drugs that treat serious conditions, and fill an unmet medical need based on a surrogate endpoint.\nA surrogate endpoint is a marker, such as a laboratory measurement, radiographic image, physical sign or other measure that is thought to predict clinical benefit but is not itself a measure of clinical benefit.\nThe use of a surrogate endpoint can considerably shorten the time required prior to receiving FDA approval.", "title": "Accelerated Approval Program - FDA", "url": "https://www.fda.gov/drugs/nda-and-bla-approvals/accelerated-approval-program", "date": "2026-05-14", "last_updated": "2026-05-18", "source": "web" }, { "id": 5, "snippet": "32 \n“21 C.F.R.\n§§ 314.500 et seq.\n& 601.40 et seq.; 57 Fed.\nReg.\n58942, 58942 (Dec. 11, 1992).\nendpoints which, while not direct measures of \nclinical benefit, were known to demonstrably\n…\n...\n33 What was new was the idea that approval \ncould rest on surrogate endpoints that were \n“reasonably likely to predict clinical benefit” but\nabout which there may be lingering uncertainty \nas to the relationship to actual clinical benefit.\n34 \n...\nbenefits and risks.36 Recognizing that reliance on \na surrogate endpoint “almost always introduces \nsome uncertainty into the risk/benefit assessment,”\nan accelerated approval sponsor was thus required \nto “persuasively support the reasonableness of \n...\nconfirmatory study.37\n...\naffected the overall risk-benefit calculation, \n...\nwas entirely appropriate, notwithstanding \nlingering uncertainty about the predictive value \nof those endpoints.\nApproval was predicated \non appropriate evidence that the endpoints in \nquestion were predictive of (i.e., “reasonably likely” \nto predict) clinical benefit, and that the risk-benefit\n…\napproval is predicated on a demonstration of \nclinical benefit, such as prolonged life or better \nquality of life, or an established, validated \nsurrogate endpoint like disease-free survival \nfor oncology drugs.45 Like traditional approval, \naccelerated approval is based, in relevant part, on \nsubstantial evidence of efficacy from adequate \nand well-controlled clinical trials.\nAn accelerated \napproval sponsor must produce data establishing \nthat the surrogate endpoint or intermediate \nclinical endpoint is “reasonably likely”46 to predict", "title": "[PDF] FDA's Accelerated Approval Pathway", "url": "https://rarediseases.org/wp-content/uploads/2021/06/NRD-2182-Policy-Report_Accelerated-Approval_FNL.pdf", "date": "2021-06-11", "last_updated": "2025-11-08", "source": "web" }, { "id": 6, "snippet": "The accelerated approval pathway authorizes the use of investigational treatments based on preliminary evidence, such as surrogate endpoints that predict clinical benefits.\n...\n1. Reliance on Surrogate Endpoints: Accelerated approval pathways are based on a product demonstrating an effect on a surrogate endpoint that is reasonably likely to predict clinical benefit, or on a clinical endpoint that can be measured earlier than irreversible morbidity or mortality.^1, 2^ Drugs are often approved based on measures that may not translate into real-world clinical benefits, especially in oncology.", "title": "What Is the Accelerated Approval Pathway? Understanding the ...", "url": "https://www.centerwatch.com/insights/what-is-the-accelerated-approval-pathway-understanding-the-latest-fda-guidance/", "date": "2025-04-14", "last_updated": "2026-05-20", "source": "web" }, { "id": 7, "snippet": "21 C.F.R.\n§ 314.510 Approval Based On a Surrogate Endpoint Or On an Effect On a Clinical Endpoint Other Than Survival Or Irreversible Morbidity", "title": "Subpart H. Accelerated Approval of New Drugs For Serious Or Life ...", "url": "https://cfr.vlex.com/source/cfr-code-federal-regulations-14106/chapter/subpart-h-accelerated-approval-of-new-drugs-for-serious-or-life-threatening-illnesses-398641958", "date": null, "last_updated": "2026-03-19", "source": "web" } ], "type": "search_results", "queries": [ "21 CFR 314 Subpart H accelerated approval surrogate endpoint definition", "FDA guidance accelerated approval confirmatory trials oncology", "21 CFR 314.500 eligibility serious or life-threatening illness surrogate endpoint requirement" ] }, { "contents": [ { "snippet": "# 21 CFR Part 314 - Subpart H - Accelerated Approval of New Drugs for Serious or Life-Threatening Illnesses\n1. § 314.500 Scope.\n2. § 314.510 Approval based on a surrogate endpoint or on an effect on a clinical endpoint other than survival or irreversible morbidity.\n3. § 314.520 Approval with restrictions to assure safe use.\n4. § 314.530 Withdrawal procedures.\n5. § 314.540 Postmarketing safety reporting.\n6. § 314.550 Promotional materials.\n7. § 314.560 Termination of requirements.\n\nSource:\n\n57 FR 58958, Dec. 11, 1992, unless otherwise noted.", "title": "21 CFR Part 314 - Subpart H | US Law", "url": "https://www.law.cornell.edu/cfr/text/21/part-314/subpart-H" }, { "snippet": "The FDA instituted its Accelerated Approval Program to allow for earlier approval of drugs that treat serious conditions, and fill an unmet medical need based on a surrogate endpoint.  A surrogate endpoint is a marker, such as a laboratory measurement, radiographic image, physical sign or other measure that is thought to predict clinical benefit but is not itself a measure of clinical benefit. The use of a surrogate endpoint can considerably shorten the time required prior to receiving FDA approval.\n\nDrug companies are still required to conduct studies to confirm the anticipated clinical benefit. If the confirmatory trial shows that the drug actually provides a clinical benefit, then the FDA grants traditional approval for the drug. If the confirmatory trial does not show that the drug provides clinical benefit, FDA has regulatory procedures in place that could lead to removing the drug from the market.\n\n## Infectious Disease Indications Accelerated Approvals\n- Ongoing (excluding vaccines)\n- Verified Clinical Benefit (excluding vaccines)\n- Ongoing (vaccines)\n- Verified Clinical Benefit (vaccines)\n- Withdrawn\n\n## Non-malignant Hematological, Neurological, and Other Disorder Indications Accelerated Approvals\n- Ongoing\n- Verified Clinical Benefit\n- Withdrawn\n\n## Cancer Accelerated Approvals\n- Ongoing\n- Verified Clinical Benefit\n- Withdrawn\n- Other", "title": "Accelerated Approval Program - FDA", "url": "https://www.fda.gov/drugs/nda-and-bla-approvals/accelerated-approval-program" } ], "type": "fetch_url_results" }, { "id": "msg_cc5071ef-5c2c-4133-8ca4-b582f1d35d07", "content": [ { "text": "The accelerated approval pathway under 21 CFR 314 Subpart H allows earlier approval of drugs for serious or life‑threatening illnesses based on effects on surrogate or intermediate clinical endpoints that are reasonably likely to predict clinical benefit, with a binding obligation to complete postmarketing confirmatory trials. [page:0][web:2][web:4] If those confirmatory trials fail to verify benefit or are not conducted with due diligence, FDA can withdraw the approval using streamlined procedures. [page:0][web:4][web:5]\n\n## Regulatory framework and scope\n\n- Subpart H (21 CFR 314.500–314.560) applies to new drugs for **serious or life‑threatening** illnesses and establishes special rules for approval on surrogate or non‑definitive clinical endpoints, use restrictions, and withdrawal. [page:0][web:2][web:5] \n- The program is intended to address **unmet medical need** by allowing earlier access while shifting some evidentiary burden (verification of clinical benefit) to the postmarketing period. [web:3][web:4][web:5]\n\nKey sections under Subpart H include: scope (§314.500), approval based on surrogate/intermediate endpoints (§314.510), approval with restrictions (§314.520), withdrawal procedures (§314.530), postmarketing safety reporting (§314.540), and promotional material requirements (§314.550). [page:0]\n\n## Eligibility criteria\n\nTo qualify for accelerated approval under 21 CFR 314 Subpart H, a product must meet both disease‑level and evidence‑level criteria. [web:2][web:4][web:5]\n\nDisease/clinical context:\n\n- The drug must treat a **serious or life‑threatening** condition (e.g., cancers, certain infections, rare progressive disorders). [web:2][web:3][web:4] \n- There must be an **unmet medical need**, such as lack of satisfactory alternatives or the drug providing a meaningful advantage over existing therapy (e.g., improved response rate where survival data are immature). [web:2][web:3]\n\nEvidence and endpoint context:\n\n- The application must show **substantial evidence of effectiveness** from adequate and well‑controlled trials, but the primary endpoint can be a surrogate or intermediate clinical endpoint instead of definitive outcomes like overall survival. [web:2][web:5][web:6] \n- FDA must find that the endpoint is **“reasonably likely to predict”** clinical benefit (e.g., effect on irreversible morbidity or mortality, or on how patients feel, function, or survive), recognizing greater residual uncertainty than in traditional approval. [web:2][web:3][web:5]\n\nFDA also considers:\n\n- Biological plausibility linking the drug’s mechanism to the surrogate or intermediate endpoint. [web:2][web:3] \n- Strength and consistency of data correlating the endpoint with clinical outcomes in the same or similar disease settings, possibly with related drugs. [web:2][web:3][web:5]\n\n## Surrogate and intermediate endpoints\n\nUnder Subpart H, accelerated approval can be based on either a surrogate endpoint or an intermediate clinical endpoint as described in §314.510. [page:0][web:2][web:3]\n\nDefinitions and characteristics:\n\n- A **surrogate endpoint** is a marker, such as a lab test, imaging measure, or physical sign, that is **thought** to predict clinical benefit but is **not itself** a direct measure of benefit (e.g., tumor shrinkage, viral load, biomarker levels). [web:2][web:4] \n- A surrogate appropriate for accelerated approval is one that is **reasonably likely to predict** clinical benefit, not one already validated to definitively predict benefit. Validated surrogates generally support traditional rather than accelerated approval. [web:2][web:3][web:5] \n- An **intermediate clinical endpoint** is a clinical measure that reflects a therapeutic effect and can be observed earlier than irreversible morbidity or mortality (e.g., progression‑free survival instead of overall survival). [web:2][web:3][web:6]\n\nRegulatory implications:\n\n- If the endpoint is already a **validated** surrogate known to predict clinical benefit (for the specific disease and intervention class), FDA typically expects the sponsor to seek traditional approval instead of using Subpart H. [web:2][web:3][web:5] \n- When evidence is insufficient to show that a biomarker or early clinical measure is either validated or reasonably likely to predict benefit, neither traditional nor accelerated approval is available. [web:2][web:3]\n\nIn practice, common examples in oncology include objective response rate or progression‑free survival used for accelerated approval, with overall survival or long‑term symptom/functional outcomes confirmed later. [web:2][web:3][web:6]\n\n## Confirmatory trial requirements\n\nAccelerated approval is **conditional** on postmarketing studies designed to verify and describe actual clinical benefit. [web:2][web:4][web:5]\n\nCore requirements:\n\n- Sponsors must conduct one or more **adequate and well‑controlled confirmatory trials** to verify that the drug truly provides clinical benefit (e.g., improved survival, reduced irreversible morbidity, or better patient function/quality of life). [web:2][web:4][web:5] \n- These trials generally use **hard clinical endpoints** rather than the surrogate or intermediate endpoint that supported approval, unless the intermediate endpoint itself becomes accepted as a direct measure of benefit. [web:2][web:3][web:5] \n- Confirmatory studies are expected to be initiated **promptly**, frequently already underway or agreed upon at the time of accelerated approval, with clear timelines and milestones. [web:2][web:3]\n\nConsequences and regulatory tools:\n\n- If postmarketing trials confirm benefit, FDA can **convert** the accelerated approval into traditional approval, effectively removing the conditional aspect. [web:4][web:5] \n- If trials **fail to verify clinical benefit**, do not show the anticipated magnitude of benefit relative to risk, or are not conducted with due diligence, FDA may use §314.530’s streamlined **withdrawal procedures** to remove the indication or product. [page:0][web:4][web:5] \n- Under §314.520, FDA may also impose **restrictions to assure safe use** (e.g., limited distribution, training or certification requirements, special monitoring) while confirmatory data are pending. [page:0][web:5]\n\nSponsors must also comply with postmarketing safety reporting (§314.540) and submit **promotional materials** for FDA review, to ensure that communications do not overstate unverified clinical benefits. [page:0][web:5]", "type": "output_text", "annotations": [], "logprobs": [] } ], "role": "assistant", "status": "completed", "type": "message" } ], "status": "completed", "error": null, "usage": { "input_tokens": 10702, "output_tokens": 1549, "total_tokens": 12251, "cost": { "currency": "USD", "input_cost": 0.00378, "output_cost": 0.01549, "total_cost": 0.02573, "cache_creation_cost": null, "cache_read_cost": 0.00096, "tool_calls_cost": 0.0055 }, "input_tokens_details": { "cache_creation_input_tokens": 0, "cache_read_input_tokens": 7680, "cached_tokens": 7680 }, "tool_calls_details": { "fetch_url": { "invocation": 1 }, "search_web": { "invocation": 1 } }, "output_tokens_details": { "reasoning_tokens": 0 } }, "background": false, "completed_at": 1779391734, "frequency_penalty": 0, "incomplete_details": null, "instructions": "## Abstract\n\nYou are an AI assistant developed by Perplexity AI. Given a user's query, your goal is to generate an expert, useful, factually correct, and contextually relevant response by leveraging available tools and conversation history. First, you will receive the tools you can call iteratively to gather the necessary knowledge for your response. You need to use these tools rather than using internal knowledge. Second, you will receive guidelines to format your response for clear and effective presentation. Third, you will receive guidelines for citation practices to maintain factual accuracy and credibility.\n\n\n## Instructions\n\nBegin each turn with tool calls to gather information. You must call at least one tool before answering, even if information exists in your knowledge base. Decompose complex user queries into discrete tool calls for accuracy and parallelization. After each tool call, assess if your output fully addresses the query and its subcomponents. Continue until the user query is resolved or until the below is reached. End your turn with a comprehensive response. Never mention tool calls in your final response as it would badly impact user experience.\n\n Make at most three tool calls before concluding.\n\n\n## Citation Instructions\n\nYour response must include at least 1 citation. Add a citation to every sentence that includes information derived from tool outputs.\nTool results are provided using `id` in the format `type:index`. `type` is the data source or context. `index` is the unique identifier per citation.\n are included below.\n\n\n- `web`: Internet sources\n- `page`: Full web page content\n- `conversation_history`: past queries and answers from your interaction with the user\n\n\n\nUse brackets to indicate citations like this: [type:index]. Commas, dashes, or alternate formats are not valid citation formats. If citing multiple sources, write each citation in a separate bracket like [web:1][web:2][web:3].\n\nCorrect: \"The Eiffel Tower is in Paris [web:3].\"\nIncorrect: \"The Eiffel Tower is in Paris [web-3].\"\n\n\nYour citations must be inline - not in a separate References or Citations section. Cite the source immediately after each sentence containing referenced information. If your response presents a markdown table with referenced information from `web`, `memory`, `attached_file`, or `calendar_event` tool result, cite appropriately within table cells directly after relevant data instead in of a new column. Do not cite `generated_image` or `generated_video` inside table cells.\n\n## Response Guidelines\n\nResponses are displayed on web interfaces where users should not need to scroll extensively. Limit responses to 5 sections maximum. Users can ask follow-up questions if they need additional detail. Prioritize the most relevant information for the initial query.\n\n### Answer Formatting\n- Begin with a direct 1-2 sentence answer to the core question.\n- Organize the rest of your answer into sections led with Markdown headers (using ##, ###) when appropriate to ensure clarity (e.g. entity definitions, biographies, and wikis).\n- Your answer should be at least 3 sentences long.\n- Each Markdown header should be concise (less than 6 words) and meaningful.\n- Markdown headers should be plain text, not numbered.\n- Between each Markdown header is a section consisting of 2-3 well-cited sentences.\n- When comparing entities with multiple dimensions, use a markdown table to show differences (instead of lists).\n- Whenever possible, present information as bullet point lists to improve readability.\n- You are allowed to bold at most one word (**example**) per paragraph. You can't bold consecutive words.\n- For grouping multiple related items, present the information with a mix of paragraphs and bullet point lists. Do not nest lists within other lists.\n\n### Tone\n\nExplain clearly using plain language. Use active voice and vary sentence structure to sound natural. Ensure smooth transitions between sentences. Avoid personal pronouns like \"I\". Keep explanations direct; use examples or metaphors only when they meaningfully clarify complex concepts that would otherwise be unclear.\n\n\n### Lists and Paragraphs\n\nUse lists for: multiple facts/recommendations, steps, features/benefits, comparisons, or biographical information.\n\nAvoid repeating content in both intro paragraphs and list items. Keep intros minimal. Either start directly with a header and list, or provide 1 sentence of context only.\n\nList formatting:\n- Use numbers when sequence matters; otherwise bullets (-) with a space after the dash.\n- Use numbers when sequence matters; otherwise bullets (-).\n- No whitespace before bullets (i.e. no indenting), one item per line.\n- Sentence capitalization; periods only for complete sentences.\n\nParagraphs:\n- Use for brief context (2-3 sentences max) or simple answers\n- Separate with blank lines\n- If exceeding 3 consecutive sentences, consider restructuring as a list\n\n\n### Summaries and Conclusions\n\nAvoid summaries and conclusions. They are not needed and are repetitive. Markdown tables are not for summaries. For comparisons, provide a table to compare, but avoid labeling it as 'Comparison/Key Table', provide a more meaningful title.\n\n\n## Prohibited Meta-Commentary\n\n- Never reference your information gathering process in your final answer.\n- Do not use phrases such as:\n- \"Based on my search results...\"\n- \"Now I have gathered comprehensive information...\"\n- \"According to my research...\"\n- \"My search revealed...\"\n- \"I found information about...\"\n- \"Let me provide a detailed answer...\"\n- \"Let me compile this information...\"\n- \"Short Answer: ...\"\n- Begin answers immediately with factual content that directly addresses the user's query.\n\n\n\n- Never reproduce copyrighted content (text, lyrics, etc.)\n- You may share public domain content (expired copyrights, traditional works)\n- When copyright status is uncertain, treat as copyrighted\n- Keep summaries brief (under 30 words) and original — don't reconstruct sources\n- Brief factual statements (names, dates, facts) are always acceptable\n\n\nCurrent date: Thursday, May 21, 2026\n\n", "max_output_tokens": 8192, "max_tool_calls": null, "metadata": {}, "parallel_tool_calls": true, "presence_penalty": 0, "previous_response_id": null, "prompt_cache_key": null, "reasoning": null, "safety_identifier": null, "service_tier": "default", "store": true, "temperature": 1, "text": { "format": { "type": "text" } }, "tool_choice": "auto", "tools": [ { "type": "web_search" }, { "type": "fetch_url" } ], "top_logprobs": 0, "top_p": 1, "truncation": "disabled", "user": null } ``` When you override a parameter, the preset's other defaults remain in effect. For example, if you override `model` on `pro-search`, you still get the `web_search` and `fetch_url` tools, the optimized system prompt, and the default reasoning steps. The full system prompts and detailed configurations for each preset are shown in the [System Prompts](#system-prompts) section above. The table at the top of this page summarizes the key parameters (model, search configuration, max steps, and available tools) for each preset. ## Frozen Configurations If you need a setup that does not change when Perplexity ships preset improvements — for example, change-managed environments, regulated workflows, or applications that need to pin a specific model and tool setup — replace the `preset` parameter with the explicit underlying configuration. This gives you the same behavior the preset has today, locked to the exact model, system prompt, and parameters you copied. To freeze a preset, copy the values from the [Available Presets](#available-presets) table and the matching system prompt from the [System Prompts](#system-prompts) section, then pass them directly instead of the preset name. See the [Using Presets](#using-presets) section above for side-by-side dynamic and frozen examples for each preset. **Dynamic vs. frozen — which to choose?** * Choose the **dynamic preset** (default) if you want the best Perplexity-recommended quality at a stable cost/latency band, and are comfortable with the underlying model or system prompt evolving over time. * Choose a **frozen configuration** if insulating your application from future preset updates matters more than picking up improvements automatically — for example, regulated workflows, change-managed environments, or contracts that require a specific underlying model and tool setup. You can mix both: use the dynamic preset in most environments, and pin a frozen configuration in places where stability is required. ## Choosing a Preset * **fast-search**: Simple questions, quick answers, minimal latency * **pro-search**: Standard queries requiring research and tool use * **deep-research**: Complex analysis, multi-step reasoning, comprehensive research * **advanced-deep-research**: Maximum depth research with institutional-grade analysis, enhanced tool access, and sophisticated source coverage ## Next Steps Get started with the Agent API. Explore direct model selection and third-party models. View complete endpoint documentation.