Skip to content

Caedral technology

Available

Caedral Rerank

Vector search returns a pile of maybe-relevant chunks. Reranking scores those candidates against the original query and puts the useful ones first. Caedral hosts the reranker; usage bills the External Models pool.

caedral-rerank · BAAI/bge-reranker-v2-m3 · $0.0005 / request

Get API keySpecialized models docs

Where it sits in a pipeline

  1. Query
  2. Retrieval
  3. Candidate documents
  4. Caedral Rerank
  5. Best results

Retrieval (often Caedral Embed) is recall-oriented. Rerank is precision-oriented. Notre does not run on this path.

What

You send a query and a list of documents. Caedral returns those documents ordered by relevance, optionally truncated with top_n. Maximum 100 documents per request.

Why

Nearest-neighbor search is fast and cheap, and it is not always the best ranking for the question. A reranker reads query and document together, which usually improves grounded generation and search quality for the same candidate set.

Use

Call POST /v1/rerank with either caedral-rerank or the canonical id. Hobby has no External Models pool — rerank requires a paid plan (or on-demand after included external usage).

  • Hosted by Caedral; billed as External Models
  • Weights are the published BAAI reranker, not a Caedral foundation model
  • Same API key as chat and embeddings

POST /v1/rerank

curl https://api.caedral.com/v1/rerank \
  -H "Authorization: Bearer $CAEDRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "caedral-rerank",
    "query": "How does Caedral billing work?",
    "documents": [
      "Pools reset monthly.",
      "On-demand accrues in $5 increments."
    ],
    "top_n": 2
  }'

Next