Skip to content
ResetData Docs

Reranking

A reranker scores how well each document answers a query. Unlike embeddings, it sees the query and the document together, which makes it considerably more accurate — and more expensive per comparison, which is why it is used to reorder a shortlist rather than search a corpus.

The usual pattern: retrieve perhaps 50 candidates with embeddings, rerank them, keep the top 5, pass those to a model.

Cohere-style — POST /rerank

Section titled “Cohere-style — POST /rerank”
Terminal window
curl https://api.resetdata.ai/api/v1/rerank \
  -H "Authorization: Bearer $RESETDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "RERANK_MODEL",
    "query": "How is usage billed?",
    "documents": [
      "Invoices are emailed with a PDF attached.",
      "Each request is charged against the workspace wallet.",
      "Passkeys can be registered from the security page."
    ],
    "top_n": 2
  }'

Returns results with an index into your original array and a relevance_score. Reorder your own list using index — don’t assume the response echoes your documents back.

NVIDIA NIM-style — POST /reranking

Section titled “NVIDIA NIM-style — POST /reranking”
Terminal window
curl https://api.resetdata.ai/api/v1/reranking \
  -H "Authorization: Bearer $RESETDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "RERANK_MODEL",
    "query": { "text": "How is usage billed?" },
    "passages": [
      { "text": "Invoices are emailed with a PDF attached." },
      { "text": "Each request is charged against the workspace wallet." }
    ]
  }'

Same idea, different envelope: the query is an object with text, and documents are objects in passages.

Practical notes

Section titled “Practical notes”
  • Scores are relative, not absolute. A model’s scores are comparable within one response, not across queries or models. Don’t hard-code a threshold from a single sample — calibrate on your own data.
  • Rerank a shortlist. Cost and latency scale with the number of documents, so retrieve broadly and rerank narrowly.
  • Mind document length. Rerankers have limited context; long documents may be truncated. Rerank the same chunks you embedded.
  • top_n saves bandwidth, not compute. Every document is still scored.