Exa for AI agents

Exa is a search engine built for software rather than people. You give it a description of what you are looking for, in a sentence, and it returns pages that mean that, ranked by meaning rather than by keyword overlap, with the page text already attached so the agent can read the sources without a second call. That one design decision is why it is the default search tool in Vaaya's network: an agent's query is almost always a description ("recent engineering write-ups on x402 facilitators"), and a keyword engine makes the agent do the translation.

What Exa does well

Semantic recall. Exa's index is built around embeddings, so a query phrased as a description finds pages a keyword engine misses, and does not need the exact words the page uses. It also has a keyword mode and an auto mode that picks per query.

Content in the same call. Ask for contents: { text: true } and every result comes back with its full clean text. For most research loops this removes the search-then-scrape chain entirely.

Filters that matter to agents. Published-date windows, include and exclude domain lists, and a category that narrows the index to news, research papers, GitHub, companies, or people.

People and companies by description. With category: "people" an open natural-language query such as "VP Sales at fintech companies with 21 to 100 employees in London" returns matching profiles. This is the fastest way in the network to turn an ideal-customer description into a candidate list.

Scale per call. Through Vaaya's REST path, numResults goes up to 100 per search, which is what makes broad sweeps economic at a cent.

Actions and prices through Vaaya

Every call is priced before it runs and billed only on success. The agent sets max_cost_cents on each call and anything above it is refused before Exa is contacted.

Example request

{
  "service": "exa",
  "action": "search",
  "params": {
    "query": "engineering write-ups on running an x402 facilitator in production",
    "type": "auto",
    "numResults": 10,
    "contents": { "text": true },
    "start_published_date": "2026-01-01"
  },
  "max_cost_cents": 2
}

Ten results with their text, one cent, a receipt on the agent's key. The same call with max_cost_cents: 0 is refused with the actual price and costs nothing.

When to use Exa

Pick Exa first for almost any "find pages about X" task. Its recall is the strongest in the network for descriptive queries, and inline content saves the second call.

Pick something else in four cases. When the answer depends on how recent a page is, Perplexity's search returns a date and last_updated per hit and filters by hour, day, or week. When you want a written answer with citations rather than a result list, Linkup's sourcedAnswer or Tavily's include_answer do that in one call. When you need a second, independent index to corroborate a result, Brave runs its own crawler and is the right racing partner. When you want the actual paper rather than a blog post about it, Valyu's academic search returns the text from arXiv and PubMed with the DOI.

Usage notes

Without a date filter, semantic search returns strong but sometimes stale results for "recent" queries. Add start_published_date whenever recency matters. The people and company categories do not support date filters. For text extraction from URLs you already hold, use exa/contents rather than searching again; it is a tenth of the price.

Exa's own documentation is at exa.ai. The search recipe that chains Exa with scraping and corroboration is at /recipes/search-scrape.

Questions

What is the Exa API good for?

Semantic web search where the query is a description of what you want rather than exact keywords: research on a topic, finding competitor content, API documentation, GitHub issues, and people or companies matching an open description. It returns full page text inline so an agent can read the sources in the same call.

How much does an Exa search cost through Vaaya?

One cent per search, with up to 100 results per call. Reading the content of URLs you already have is a tenth of a cent per URL per content field. Billed only on success; the agent states a maximum cost on every call.

When should an agent use Tavily or Brave instead of Exa?

Tavily when you want a synthesized answer with the results. Brave when you want a second, independent keyword index to corroborate a semantic result, or news with freshness filters. Exa is the default for everything else.