Send query and documents
First obtain an inexpensive broad set through search or embeddings, then send query and candidates to rerank. This reduces cost and latency compared with reranking a whole corpus.
Documents need to retain your internal id so a result can map back to the source record.
Rerank request
bash
curl --request POST https://cicora.ai/api/v1/rerank \
--header "Authorization: Bearer $CICORA_API_KEY" \
--header 'Content-Type: application/json' \
--data '{
"model": "cohere/rerank-v3.5",
"query": "How do I change a subscription?",
"documents": ["Billing changes are in Settings.", "Images can be generated from a prompt."],
"top_n": 1
}'Candidate size
- Limit documents to a sensible count before the call.
- top_n should reflect how many items the next step really needs.
- Do not combine documents with different access rights into one text.
Use the score
A rerank result is an ordering signal, not absolute truth. Combine it with access filters, recency, and business rules.
Log the query and de-identified candidate ids for offline evaluation without copying restricted content into unnecessary logs.