Universal Reranker (flow) icon

Universal Reranker (flow)

Rerank documents in workflow using various services (vLLM, LocalAI, Infinity, Cohere)

Universal Reranker (flow)

AI-generated

Summary

Rerank documents against a query using an OpenAI‑compatible endpoint or the Cohere rerank API, returning the top‑K results that meet a relevance threshold.

Inputs

  • query (required) — The search query to rerank documents against.
  • documentsField (required) — Field name on each input item that holds the array of documents to rerank.
  • service (required) — Reranking service to use: openai-compatible or cohere.
  • endpoint — URL of the OpenAI‑compatible reranker endpoint (shown when service is openai-compatible).
  • model — Model identifier for the OpenAI‑compatible reranker (shown when service is openai-compatible).
  • enableCustomTemplates — Whether to use custom query/document templates for special models (shown when service is openai-compatible).
  • templatePreset — Template preset to use when custom templates are enabled (shown when service is openai-compatible and enableCustomTemplates is true).
  • instruction — Instruction used by the Qwen3 template (shown when service is openai-compatible, enableCustomTemplates is true and templatePreset is qwen3).
  • queryPrefix — Text prepended to the query when using custom templates (shown when service is openai-compatible, enableCustomTemplates is true and templatePreset is custom).
  • querySuffix — Text appended to the query when using custom templates (shown when service is openai-compatible, enableCustomTemplates is true and templatePreset is custom).
  • documentPrefix — Text prepended to each document when using custom templates (shown when service is openai-compatible, enableCustomTemplates is true and templatePreset is custom).
  • documentSuffix — Text appended to each document when using custom templates (shown when service is openai-compatible, enableCustomTemplates is true and templatePreset is custom).
  • cohereModel — Cohere model to use (shown when service is cohere).
  • cohereCustomModel — Custom Cohere model name when cohereModel is set to custom (shown when service is cohere).
  • topK (required) — Maximum number of documents to return after reranking.
  • threshold (required) — Minimum relevance score (0‑1) a document must have to be included in the output.
  • includeOriginalScores (required) — Whether to keep the original document scores in the output.
  • enableCache (required) — Whether to cache reranking results to speed up repeated queries.
  • cacheTtl — Time‑to‑live for cached results in minutes (shown when enableCache is true).

Output shape

Each input item is returned with the following additional fields:

All original input fields are preserved; the node adds these properties: rerankedDocs (array of reranked documents), originalCount (number of documents before reranking), rerankedCount (number after applying topK and threshold), query (the query used for this operation).

Examples

Example 1: Rerank a list of search snippets against a user question using a local vLLM endpoint.

Set service to openai-compatible, endpoint to http://localhost:8000/v1/rerank, model to BAAI/bge-reranker-v2-m3, topK to 5, threshold to 0.3.

Example 2: Use Cohere’s multilingual reranker to reorder product descriptions.

Set service to cohere, cohereModel to rerank-multilingual-v3.0, provide your Cohere API key via credentials, topK to 10, threshold to 0.

Discussion