Exa Deep Search

Exa Deep Search

Turn Exa search into a focused, source-backed research brief. This Skill calls the named Exa capabilities in the SandBase API map through the SandBase MCP gateway. In a SandBase Agent, run the capabilities directly. In another compatible agent, require an authorized SandBase connection before starting; never request, print, or store an API key in the research output.

Read example workflows when the user needs a starting prompt or wants to understand the output.

Operating principles

• Start from the user's research question and decision context, not a generic search.
• Treat Exa results as evidence; treat model-generated synthesis, comparisons, and recommendations as judgment clearly separated from sources.
• Select search depth, time window, domains, and geography deliberately. State any assumption rather than silently defaulting.
• Optimize for source quality, recency, and relevance — not quantity.
• Cite every externally verifiable claim with a result URL and publication date (when available).
• Keep user research goals, company context, and strategy confidential unless sharing is explicitly requested.

Workflow

1. Frame the research question

Collect or infer: the topic or entity, time window, geography, trusted or excluded domains, audience for the deliverable, and how findings will be used. Classify the request as one or more of: landscape scan, deep evidence gathering, competitive intelligence, current news monitoring, or specific-source extraction.

When the research question is broad, propose 2–3 focused sub-queries and confirm scope before spending API calls.

2. Select and call SandBase capabilities

Read the SandBase API map before selecting tools. Use the listed tool_name through the SandBase gateway:

1. Call sandbase_describe_tool for the selected tool_name and read its current input schema.
2. Call sandbase_call_tool with that exact tool_name and only schema-defined arguments.
3. Keep the tool name, query, search parameters, and result metadata with the returned data.

3. Search with Exa

Use exa_search with parameters matched to the research need:

Tips:

• Write queries as natural-language statements of what a good result page would say, not short keyword strings. Exa responds best to semantic queries.
• Use category when available (e.g., "research paper", "company", "news") to narrow result types.
• Iterate: refine by entity, product, problem, event, or time period until evidence is sufficient.
• Request include_highlights: true to get relevant snippets without extracting full text for every result.

4. Extract selected sources

When deeper analysis of specific pages is needed, send selected URLs to exa_contents:

• Choose include_text: true for full page content when analyzing structure or extracting data.
• Choose include_highlights: true with a highlights_query to focus extraction on specific aspects.
• Choose include_summary: true for concise overviews when reviewing many pages.
• Use subpages only for explicit documentation, pricing, or API crawl tasks.
• Use max_age_hours: 0 only when freshness requires a live crawl; avoid for routine research.

If exa_contents is not yet available in the current Gateway, return the Search results and explicitly state that extraction is awaiting capability publication.

5. Synthesize findings

• Separate direct observations from interpretation.
• Group findings by theme, entity, or chronology as appropriate for the research question.
• Note disagreements between sources and evidence gaps.
• Propose follow-up queries for unresolved questions.

Query crafting tips

Good Exa queries describe the content of the ideal result page:

• Add temporal context: "in 2025", "since January", "latest announcement".
• Add specificity: mention the industry, company size, technology stack, or use case.
• Use exclude_domains to avoid results you already know about.

Output

Return a structured research brief:

Source map

Key findings

Numbered findings, each citing source(s) by number.

Disagreements and evidence gaps

What sources disagree on, and what questions remain unanswered.

Suggested next queries

Follow-up Exa queries or alternative research paths.

Evidence rules

• Cite a result URL for every externally verifiable claim.
• Label a result's publication date as "unavailable" when Exa does not return one.
• Do not treat an Exa summary as a source quote; use it as an aid to select evidence, then cite the original URL.
• Do not call Exa Answer or Exa Agent endpoints. The user's Agent/LLM synthesizes the evidence.
• Do not copy long source passages; paraphrase and cite.
• Mark clearly when a finding is inferred from multiple sources vs. directly stated in one.

Failure handling

• If SandBase is unavailable or unauthorized, report the failed capability and ask the user to connect or authorize SandBase; do not silently substitute a direct provider API.
• If exa_search returns few or no results, try: broader query, different search_depth, removed domain filters, or a wider date range. Report if the topic genuinely lacks public coverage.
• If exa_contents is unavailable, deliver search results with highlights and explicitly note the extraction gap.
• If results are low-quality or off-topic, refine the query before reporting; explain what was tried.

Example tasks

• "Find the last 30 days of reliable sources about AI agent observability. Give me a five-source brief with gaps."
• "Research how enterprise teams evaluate AI agents. Prefer company and academic sources; exclude vendor blogs."
• "Compare the public arguments for and against a retrieval architecture. Use advanced search and cite each source."
• "Find recent funding announcements in the AI developer tools space. Only include sources from the last 7 days."
• "Extract the pricing and feature comparison from these three competitor pages: [URLs]."

Quality gate

Before delivering, verify that:

• Every finding cites at least one source URL.
• Observations are separated from model-generated interpretations.
• The search parameters (depth, dates, domains) match the stated research need.
• Evidence gaps and low-confidence findings are explicitly labeled.
• The deliverable format matches what the user requested.