Models and Modes Reference
Stable selection guidance for Exa search types and compatibility-layer model routing.
Canonical Docs Links
- Base docs URL:
https://exa.ai/docs - Search reference:
/reference/search - OpenAI SDK compatibility:
/integrations/openai-sdk
Search Types
Exa's primary search modes for new work:
| Type | Use When | Tradeoff |
|---|---|---|
auto |
You want the default; omit type entirely rather than sending it |
Best general-purpose starting point |
lite |
Low-latency, low-cost chat, voice, autocomplete, and agent loops | Cached retrieval with a reduced parameter set |
deep-lite |
You want lightweight synthesis | More reasoning than default modes |
deep |
You need deeper synthesized retrieval, or a wide outputSchema whose fields take more than one search to fill |
Higher latency; runs several searches, so more schema fields come back filled |
With type: "lite", supported parameters are query (a single string), objective, userIdentifier, metadata, category, numResults, contents.highlights, contents.text, contents.snapshotAsOf, userLocation, includeDomains, excludeDomains, startPublishedDate, endPublishedDate, moderation, and compliance. Other documented search parameters are rejected, including outputSchema, systemPrompt, stream, and freshness options such as contents.maxAgeHours. Lite serves cached content without fetching pages at query time.
How Output Controls Affect Behavior
outputSchemaadds synthesis work and latency on supported search types. Lite does not support itsystemPromptguides synthesis behavior but does not replace schema controlmaxAgeHours: 0forces fresh crawling and can increase latency substantially
Do not choose a deep variant only because you want structured output. Choose it when the retrieval task itself needs more reasoning or multi-step synthesis. A wide schema whose fields take more than one search to fill is one such task: deep runs several searches, so more of the fields come back filled. A compact schema (author and URL per article) stays on auto.
If you are optimizing for responsiveness, start with:
type: "lite"when latency or cost matters more than maximum quality, or notypeat all (defaultauto) otherwise- no
outputSchema - default freshness behavior
Then add deeper reasoning or stronger freshness only when the use case requires it.
Answer and Compatibility Models
In OpenAI-compatible flows, current model routing is:
| Model | Typical Use |
|---|---|
exa |
Answer-style compatibility via /chat/completions |
exa-agent |
Exa Agent runs via /responses |
These are compatibility-layer names. Native Exa endpoint choice remains the more important design decision.
deep-reasoning is legacy. Use deep instead, or the Agent API (agent.md) for long-running research, list building and multi-hop enrichment.
Practical Defaults
auto is almost always the right type, including for coding and agent workflows. Deviate only for a real latency or reasoning requirement:
- General retrieval: omit
type; the server defaults toauto - Latency-sensitive path or real-time UX:
type: "lite"withhighlights