Sign Up
Web Search

Exa

Neural web search built for AI agents. Use Exa in your Vellum assistant for meaning-based retrieval that returns the most relevant passages from each page alongside the link.

What it does

Exa is a search engine designed for LLMs rather than people typing keywords. It embeds your query and matches it against the meaning of pages, so descriptive, natural-language questions work as well as short keyword queries. Each result carries the page title, URL, publish date when known, and highlights: the passages from the page that best answer the query.

When configured, Vellum sends web search calls to Exa, asks for highlights on every result, and feeds them back to your assistant in context. The model can cite the passages directly, open a source with web fetch, or synthesize an answer across several results.

Prerequisites

  • A running Vellum assistant. Cloud, self-hosted, or the desktop apps all work.
  • An Exa API key. Create one at dashboard.exa.ai. Exa keys are opaque strings with no fixed prefix.

Setup

Exa is BYOK (bring your own key). You stay in control of the key and pay Exa directly for usage.

  1. Open assistant settings. In Vellum, head to Settings → Models & Services.
  2. Add your Exa API key. Under the Web Search section, choose Exa as the provider and paste your API key. Vellum stores the key in your local secure store and never writes it to disk in plaintext.
  3. Try it. Start a new conversation and ask a question that needs fresh or hard-to-find information. Vellum calls Exa under the hood and feeds the highlighted passages back to the model in context.

Prefer the CLI? From any shell where the assistant daemon is running:

assistant keys set exa <your-exa-api-key>
assistant config set services.web-search.provider exa

Swap or revoke the key at any time with assistant keys delete exa.

Best practices

  • Describe what you want, not just keywords. Exa matches on meaning, so "a blog post explaining how vector databases handle deletes" retrieves better than "vector db delete". Your assistant already phrases queries this way; you rarely need to coach it.
  • Recency filters are hard cutoffs. When the assistant asks for results from the past day, week, month, or year, Vellum passes that window to Exa as a publish-date filter. Pages without a detectable publish date are dropped from filtered searches, so leave the filter off for evergreen topics.
  • Fallback behavior. Exa participates in the web search fallback chain after the other BYOK providers. The chain skips any provider without a key connected; if none are connected, the search returns an error. See the Web Search page for the full fallback rules.
  • Billing. Exa usage is billed per request directly by Exa under the account that owns the key, separately from Vellum credits.
  • Privacy. Search queries leave your assistant and reach Exa servers. Review the Exa privacy policy for details.

Resources