Skip to main content

Reranking

还没翻译译自英文（tanstack.com）本页 51 段都还没译

You have a query and a list of candidate documents — chunks from a vector search, rows from a keyword query, FAQ entries — and you need them ordered by how well they actually answer the query. Vector similarity gets you close, but a dedicated reranking model is far more precise. By the end of this guide you'll have that list reordered, with a relevance score per document.

rerank() is the precision step in a retrieval pipeline: retrieve a broad set of candidates cheaply, then rerank to surface the few that matter.

Providers

Reranking is available from two adapters today:

  • Cohere (@tanstack/ai-cohere) — cohereRerank('rerank-v3.5'), talking to Cohere directly.
  • OpenRouter (@tanstack/ai-openrouter) — openRouterRerank('cohere/rerank-v3.5'), routing rerank through your existing OpenRouter key.

Both implement the same rerank() activity — swap the adapter, keep the call.

Installation

Cohere:

OpenRouter:

Peer dependency:

Basic Usage

Pass a query and an array of documents. The result's rerankedDocuments are ordered most-relevant first, and ranking carries the relevance score and the original index of each.

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

const { ranking, rerankedDocuments } = await rerank({
  adapter: cohereRerank('rerank-v3.5'),
  query: 'talk about rain',
  documents: ['sunny day at the beach', 'rainy afternoon in the city'],
  topN: 2,
})

console.log(rerankedDocuments[0]) // 'rainy afternoon in the city'
console.log(ranking[0]) // { index: 1, score: 0.98, document: 'rainy afternoon in the city' }

The adapter reads COHERE_API_KEY from the environment. To pass a key explicitly, use createCohereRerank('rerank-v3.5', 'co-...').

To rerank through OpenRouter instead, swap the adapter — everything else stays the same:

import { rerank } from '@tanstack/ai'
import { openRouterRerank } from '@tanstack/ai-openrouter'

const { rerankedDocuments } = await rerank({
  adapter: openRouterRerank('cohere/rerank-v3.5'),
  query: 'talk about rain',
  documents: ['sunny day at the beach', 'rainy afternoon in the city'],
  topN: 2,
})

console.log(rerankedDocuments[0]) // 'rainy afternoon in the city'

openRouterRerank reads OPENROUTER_API_KEY from the environment.

Reranking Object Documents

Documents don't have to be strings. Pass JSON-serializable objects and the original object is returned in the result — fully typed — so you can carry an id or metadata through the rerank and read it back off the ranked results.

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

const chunks = [
  { id: 'doc-1', text: 'A heavy gaming desktop with an RTX card.' },
  { id: 'doc-2', text: 'A lightweight ultrabook with all-day battery.' },
]

const { ranking } = await rerank({
  adapter: cohereRerank('rerank-v3.5'),
  query: 'best laptop for travel',
  documents: chunks,
})

// `document` is the original object — `id` is available and type-safe.
console.log(ranking[0]?.document.id) // 'doc-2'

Object documents are serialized to JSON before being sent to the provider; the ranking is mapped back to your original elements by index.

Options

OptionTypeDescription
adapterRerankAdapterA rerank adapter created with a model (e.g. cohereRerank('rerank-v3.5'))
querystringThe search query documents are scored against — required
documentsArray<string | object>Candidate documents to rerank — required
topNnumberReturn only the top N results
abortSignalAbortSignalCancel the in-flight request
modelOptionsprovider optionsProvider-specific options (see below)
middlewareArray<GenerationMiddleware>Observe-only lifecycle hooks (usage, finish, error, abort)

Provider Options

Cohere rerank accepts:

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

const { ranking } = await rerank({
  adapter: cohereRerank('rerank-v3.5'),
  query: 'refund policy',
  documents: ['Returns accepted within 30 days.', 'Free shipping over $50.'],
  modelOptions: {
    // Cap tokens kept per document when chunking long inputs (Cohere default: 4096).
    maxTokensPerDoc: 512,
  },
})

console.log(ranking)

Result Shape

import type { TokenUsage } from '@tanstack/ai'

interface RerankResult<TDocument = string> {
  id: string
  model: string
  // Scored results, most relevant first.
  ranking: Array<{ index: number; score: number; document: TDocument }>
  // The documents reordered by relevance (ranking.map(r => r.document)).
  rerankedDocuments: Array<TDocument>
  // Rerank typically bills in provider "search units"
  // (usage.billed = { quantity, unit: 'units' }). Some providers (for example
  // OpenRouter) also report totalTokens and cost. Cohere reports only search
  // units and leaves token counts at 0.
  usage: TokenUsage
}

Server Endpoint

Reranking runs on the server (it needs your API key). Wrap it in an API route and call it from the client over fetch:

// routes/api/rerank.ts
import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'
import { createFileRoute } from '@tanstack/react-router'

export const Route = createFileRoute('/api/rerank')({
  server: {
    handlers: {
      POST: async ({ request }) => {
        const body: unknown = await request.json()
        if (
          typeof body !== 'object' ||
          body === null ||
          !('query' in body) ||
          typeof body.query !== 'string' ||
          !('documents' in body) ||
          !Array.isArray(body.documents)
        ) {
          return new Response('Invalid request body', { status: 400 })
        }
        const { query, documents } = body
        const topN = 'topN' in body && typeof body.topN === 'number'
          ? body.topN
          : undefined

        const result = await rerank({
          adapter: cohereRerank('rerank-v3.5'),
          query,
          documents,
          topN,
        })

        return Response.json(result)
      },
    },
  },
})
// client.ts — call the endpoint and use the reordered documents
async function rerankDocuments(query: string, documents: Array<string>) {
  const res = await fetch('/api/rerank', {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify({ query, documents, topN: 3 }),
  })
  const result = await res.json()
  return result.rerankedDocuments
}

In a RAG Pipeline

Reranking shines as the second stage after a cheap, broad retrieval. Over-fetch candidates with vector search, then rerank to keep only the most relevant few for the prompt:

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'
import { vectorSearch } from './my-vector-store'

async function retrieveContext(query: string) {
  // 1. Over-fetch candidates cheaply.
  const candidates = await vectorSearch(query, { limit: 50 })

  // 2. Rerank and keep the most relevant handful for the prompt.
  const { rerankedDocuments } = await rerank({
    adapter: cohereRerank('rerank-v3.5'),
    query,
    documents: candidates.map((c) => c.text),
    topN: 5,
  })

  return rerankedDocuments
}

Cancellation

Pass an abortSignal to cancel an in-flight request:

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

const controller = new AbortController()
setTimeout(() => controller.abort(), 5000)

const result = await rerank({
  adapter: cohereRerank('rerank-v3.5'),
  query: 'q',
  documents: ['a', 'b'],
  abortSignal: controller.signal,
})

console.log(result.rerankedDocuments)

Observability

Attach observe-only middleware to track usage, completion, errors, and cancellation — the same GenerationMiddleware contract the media activities use:

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

const result = await rerank({
  adapter: cohereRerank('rerank-v3.5'),
  query: 'q',
  documents: ['a', 'b'],
  middleware: [
    {
      name: 'usage-logger',
      onUsage: (_ctx, usage) => {
        if (usage.billed) {
          console.log(
            `search units billed: ${usage.billed.quantity} ${usage.billed.unit}`,
          )
        }
      },
    },
  ],
})

console.log(result.rerankedDocuments)

Tip: Pass otelMiddleware() to emit OpenTelemetry spans for rerank calls. See OpenTelemetry.

Environment Variables

The Cohere rerank adapter uses:

  • COHERE_API_KEY: Your Cohere API key

Error Handling

import { rerank } from '@tanstack/ai'
import { cohereRerank } from '@tanstack/ai-cohere'

try {
  const result = await rerank({
    adapter: cohereRerank('rerank-v3.5'),
    query: 'q',
    documents: ['a', 'b'],
  })
  console.log(result.rerankedDocuments)
} catch (error) {
  if (error instanceof Error) {
    console.error('Rerank failed:', error.message)
  }
}

Passing an empty documents array throws before any request is made.

Runnable Example

examples/ts-react-rerank is a small TanStack Start app that runs everything on this page: a fixed corpus of support articles listed newest-first, a query box, and a side-by-side view of the original order against the reranked order with scores. The provider dropdown switches between the Cohere and OpenRouter adapters over the same rerank() call.

cd examples/ts-react-rerank
pnpm install
cp .env.example .env   # add COHERE_API_KEY and/or OPENROUTER_API_KEY
pnpm dev

Next Steps

Edit on GitHub
AI 翻译查看英文原文