Exa

Exa searches by meaning rather than keywords. It ranks the web semantically and scores every result, so it shines on the queries where the words you would search for are not the words on the page: “papers arguing against scaling laws” finds the right papers even when none of them contain that phrase. The type knob picks the retrieval strategy and category narrows to a content vertical (news, research papers, people, and so on).

Install

pnpm add @effect-uai/core @effect-uai/exa effect

Wire it up

import { Config, Effect, Layer } from "effect"
import { FetchHttpClient } from "effect/unstable/http"
import { layer as exaLayer } from "@effect-uai/exa/ExaSearch"

const provider = Layer.unwrap(
  Effect.gen(function* () {
    const apiKey = yield* Config.redacted("EXA_API_KEY")
    return exaLayer({ apiKey })
  }),
)

const mainLayer = provider.pipe(Layer.provide(FetchHttpClient.layer))

exaLayer registers two service tags from one implementation:

ExaSearch: the typed tag. Yield this for the Exa-specific knobs below and the per-result score.
WebSearch: the generic tag for provider-portable code, including webSearchTool.

Calling it

import { ExaSearch } from "@effect-uai/exa/ExaSearch"

const program = Effect.gen(function* () {
  const search = yield* ExaSearch
  return search.search({
    query: "papers questioning scaling laws",
    maxResults: 10,
    type: "auto", // retrieval strategy
    category: "research paper", // content vertical
  })
})

Every cross-provider field from CommonSearchRequest works here. The typed request adds:

Field	Values	Meaning
`type`	`auto` `fast` `instant` `deep-lite` `deep` `deep-reasoning` …	Retrieval strategy. `auto` picks per query.
`category`	`news` `research paper` `company` `people` …	Narrow the index to a vertical.

Exa ranks, so each result carries a score. Pure search returns no snippet until you fetch page contents, which is a separate, planned capability. See what web search is not.

Exa

Install

Wire it up

Calling it

See also