Reranking

Use the standalone reranking endpoint when you already have candidate text and need AIVAX to order it against a query. It does not search a collection, create embeddings, or retrieve additional documents. It applies the local lexical reranker to the exact strings supplied by the caller.

Use Semantic Search instead when AIVAX must find candidates inside a RAG collection. Semantic Search can then apply rrf, lexical, or smart reranking after vector retrieval.

Endpoint

POST /api/v1/rank

Request limits

Limit	Value
Documents per request	1,024
Query plus all document text	32,768 characters
top_n	1 through the number of supplied documents

query must be a non-empty string. documents must be a non-empty array of strings. When top_n is omitted, the endpoint returns a score for every supplied document.

The lexical reranker considers exact tokens, fuzzy token matches, term coverage, term proximity, and document length. Results contain the original zero-based index and a normalized relevanceScore; higher scores indicate stronger lexical relevance. The document text is not repeated in the response, so map each result index back to the original documents array.

Billing and shared quota

The endpoint records a base usage price of US$0.0001 per request, equivalent to US$1 per 10,000 searches. The account's plan commission multiplier is applied when usage is recorded. This price belongs only to /api/v1/rank; using a local reranker inside Semantic Search does not add this standalone endpoint charge.

Reranking has a shared per-account limit across the standalone endpoint and every non-none reranker used by RAG search.

A Semantic Search request with reranker: "none" does not consume the reranking quota. Other RAG search limits, public-key limits, and model limits remain independent.

Reranking improves ordering only among supplied candidates. If the relevant text is missing from the input, reranking cannot recover it; improve candidate retrieval or use Semantic Search first.