exa-search

Exa Search Skill

Trigger Conditions & Endpoint Selection

Choose Exa endpoint based on user intent:

• search: Need semantic search / find web pages / research topics. Use type: "auto" by default.
• deep search / structured research: Use the search endpoint with type: "deep" or type: "deep-reasoning" and optional outputSchema.
• contents: Given result IDs, need to extract full content.
• findsimilar: Given URL, need to find similar pages.
• answer: Need direct answer to a question.

/research and /research/v1 are deprecated and were hard-removed on 2026-05-01. Do not use them for new calls; migrate research-style requests to /search with type: "deep-reasoning".

Recommended Architecture (Main Skill + Sub-skill)

This skill uses a two-phase architecture:

1. Main skill (current context): Understand user question → Choose endpoint → Assemble JSON payload
2. Sub-skill (fork context): Only responsible for HTTP call execution, avoiding conversation history token waste

Execution Method

Use Task tool to invoke exa-fetcher sub-skill, passing command and JSON (stdin):

Task parameters:
- subagent_type: Bash
- description: "Call Exa API"
- prompt: cat <<'JSON' | node scripts/exa-api.cjs <search|contents|findsimilar|answer>
  { ...payload... }
  JSON

The script still accepts the legacy research command for backwards compatibility, but it normalizes the payload and sends it to /search with type: "deep-reasoning".

Payload Examples

1) Search

cat <<'JSON' | node scripts/exa-api.cjs search
{
  "query": "Latest research in LLMs",
  "type": "auto",
  "numResults": 10,
  "category": "research paper",
  "includeDomains": [],
  "excludeDomains": [],
  "startPublishedDate": "2025-01-01",
  "endPublishedDate": "2025-12-31",
  "contents": {
    "highlights": true,
    "summary": true
  }
}
JSON

Search Types:

• auto: Balanced default
• fast: Low latency
• instant: Lowest latency
• deep-lite: Lightweight synthesized output
• deep: Multi-step search with reasoning and structured outputs
• deep-reasoning: Highest-effort deep search for complex research tasks