Deep Research

Tool Routing

Choose the right tool based on the query content:

1. URL provided → crawling_exa

If the query contains a URL (starts with http:// or https://), use crawling_exa to extract the page content. Set maxCharacters to 10000 for full extraction. Then summarize the content for the user.

2. Company name query → company_research_exa

If the query is specifically about a company (e.g., "research Stripe", "tell me about OpenAI"), use company_research_exa with the company name.

3. Programming/code question → get_code_context_exa

If the query is about code, APIs, libraries, SDKs, or programming patterns, use get_code_context_exa. Set tokensNum to 8000 for thorough results.

4. Quick factual lookup → web_search_exa

If the query is a simple factual question (who, what, when, where) that doesn't need synthesis, use web_search_exa for speed.

5. Focused question needing citations → deep_search_exa

If the query needs multi-angle search but not a full report, use deep_search_exa:

• Default type: deep
• Use deep-reasoning for queries that need analysis (triggered by pro: prefix or complex phrasing)

6. Complex research (default) → deep_researcher_start + deep_researcher_check

For anything that needs a comprehensive report with synthesis — this is the default path for most research queries.

Steps:

1. Call deep_researcher_start with:
   • instructions: The full research query, enriched with specificity if the user's query is vague
   • model: Based on the tier prefix (see table above)
2. Tell the user: "Starting Exa deep research (model: {model})... This will take {expected time}."
3. Poll by calling deep_researcher_check with the returned researchId:
   • If status is NOT completed, wait 5 seconds, then call again
   • Repeat until status is completed
   • Do NOT call more than 30 times (timeout safety)
4. Present the completed report to the user

Output Format

When presenting results:

1. Lead with the answer — synthesize the key findings in 2-3 paragraphs
2. Citations — list sources with titles and URLs
3. Offer to save — if the report is longer than ~500 words, ask: "Want me to save this report to a file?"

Examples

/exa-deep-research what is retrieval augmented generation
→ Routes to deep_researcher_start with exa-research-fast

/exa-deep-research pro: comparative analysis of transformer architectures for education
→ Routes to deep_researcher_start with exa-research-pro

/exa-deep-research https://arxiv.org/abs/2401.12345
→ Routes to crawling_exa

/exa-deep-research how does the Anthropic Python SDK handle streaming
→ Routes to get_code_context_exa

/exa-deep-research Notion company
→ Routes to company_research_exa

/exa-deep-research who won the 2024 Nobel Prize in Physics
→ Routes to web_search_exa