Search | Anara API

POST /v1/search runs a metadata search over the corpus: titles, abstracts, authors, venues. It needs the papers scope. Instant costs $2, standard $5, and deep $12 per 1,000 queries. The number of results does not change the price.

Request

field	type	notes
`query`	`string`	Natural language or keywords. 1 to 1,000 characters.
`type`	`"instant" \| "standard" \| "deep"`	`instant` is vector-only and fastest. `standard` (default) is hybrid keyword plus vector. `deep` searches open-access full text with citable passages.
`maxResults`	`int`	Default 10, max 100.
`filters`	object	See below.
`sort`	`"relevance" \| "recent" \| "citations"`	`relevance` (default) ranks by fit to the query, breaking near-ties toward papers cited unusually often for their age. `recent` returns the most relevant papers from the last two years, widening to four when few match, and demotes repository deposits. `citations` orders the relevant papers by citation count. Works with every `type`.
`includeTotal`	`boolean`	Best effort: include `totalEstimate` only if it is ready when results arrive. Prefer `/v1/search/total`.

Filters

field	type	example
`publicationYear`	`string`	`"2020"`, `">=2018"`, `"2015..2020"`
`isOpenAccess`	`boolean`	`true`
`hasFulltext`	`boolean`	`true` restricts to papers that `deep` can read
`field`	`string`	`"Computer Science"`
`citedByCount`	`string`	`">100"`

curl -X POST https://api.anara.com/v1/search \
  -H "Authorization: Bearer $ANARA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"query": "CRISPR off-target detection", "maxResults": 20, "filters": {"publicationYear": ">=2021", "hasFulltext": true}}'

Response

{
  "papers": [
    {
      "id": "W4210000000",
      "title": "…",
      "year": 2022,
      "authors": ["…"],
      "abstract": "…",
      "citedByCount": 41,
      "openAccess": true,
      "hasFulltext": true
    }
  ],
  "totalApprox": 20,

  "charge": { "costMicroUsd": 5000, "balanceMicroUsd": 9990000 }
}

totalApprox is the number of papers in this response. It is not a count of the whole corpus.

totalEstimate appears only when the request sets includeTotal and the estimate is ready by the time results are. Search never waits for it. To always get the count, call /v1/search/total alongside the search.

Total matches

POST /v1/search/total estimates how many papers in the corpus have every word of the query in their title or abstract, under the given filters, rounded to two significant figures, like the hit count a search engine shows. Counts under 1,000 are exact. Send it in parallel with /v1/search, with the same query, type and filters, so results never wait for the count. It is free and needs the papers scope.

field	type	notes
`query`	`string`	The same query as the search.
`type`	`"instant" \| "standard"`	Default `standard`. Deep has no count.
`filters`	object	The same filters as the search.

The response is { "totalEstimate": 19000 }. totalEstimate is null when the query has no words to count, for example only stopwords. It usually arrives in under a second.

Deep search

Use POST /v1/search with "type": "deep", or select Deep under Type in the playground. Deep search returns citable passages from open-access full text. Deep returns totalApprox without a corpus estimate.

Result count and coverage

The response contains up to maxResults ranked papers. Call /v1/search/total for the estimated match count.

Passages and hosted PDF links are returned only for open-access papers. A missing PDF link means no hosted copy is available; the API does not fall back to a publisher download.

Additional filters include authorName, sourceName, field, citedByCount, publicationYear, isOpenAccess, and hasFulltext. Full-text-only searches are restricted to open access.

Workspace requests are limited to 120 per minute across API keys. The signed-in playground has a separate limit of 30 per minute. A 429 response means retry after 60 seconds.

Each paper includes venue, publicationDate (the full indexed date, not just year), field, language, arxivId, pmcid, and pmid alongside doi when available. Missing metadata is omitted. openAlexId is not exposed.