Skip to main content
Exa is available as an optional web search backend in FIM One. When you enable it, agents get access to the full Exa search surface — auto (Exa’s recommended search type), neural, fast, deep-lite, deep, deep-reasoning, and instant — plus structured content retrieval (highlights, text, summaries) and rich filtering (categories, include/exclude domains, date ranges, and max-age windows). Sensible Exa-side defaults ship out of the box, so you get Exa best practices without manual tuning.

​
Why Exa in FIM One

  • Neural search for agent reasoning — Exa’s embedding-native search returns semantically relevant results, not just keyword matches. This matters for research agents that reason over web content.
  • Deep-reasoning mode for hard queries — single-call multi-hop reasoning over the open web, without the agent orchestrating sub-queries itself.
  • Built-in content extraction — highlights + text + optional summaries arrive with the search call, eliminating a second round-trip to fetch and clean page content.
  • Tight domain / recency controls — includes/excludes domain lists (up to 1200 entries), ISO-8601 date windows, and maxAgeHours for time-sensitive workflows.

​
Quick start

​
1. Get an Exa API key

Sign up at exa.ai and grab your API key from the dashboard.

​
2. Enable Exa in FIM One

Add to your .env:
Set both variables so web_search routes through Exa. Prefer an explicit WEB_SEARCH_PROVIDER=exa over relying on key auto-detect alone.

​
3. Use it from any agent

Web search is exposed to agents as a built-in tool — no extra wiring needed. With Exa configured as above, agents that have web_search enabled route search calls through Exa.

​
Configuration

All Exa-specific knobs have sensible defaults. Override any of them via environment variables:

​
How snippets are assembled

FIM One cascades through Exa’s content fields to produce a useful snippet on every result, regardless of which content types the API populated:
  1. highlights (joined with …) — preferred when present
  2. text — falls back to extracted article text, truncated to EXA_TEXT_MAX_CHARS
  3. summary — used when both above are empty
This guarantees downstream agents always have a non-empty snippet to reason over, even for results where Exa couldn’t extract highlights.

​
Preset: news monitoring

​
Preset: research paper retrieval

​
Preset: deep-reasoning agent

​
Attribution

Every Exa request from FIM One carries an x-exa-integration: fim-one header so Exa can attribute API usage back to this integration in maintainer dashboards. This is used only for integration-level attribution — no end-user identification or tracking.

​
Source

The Exa connector lives at src/fim_one/core/web/search/exa.py and follows the same BaseWebSearch protocol as Jina, Tavily, and Brave — switching providers is a single env-var change.