Reranker

Reranking is an optional post-retrieval step that re-scores results using a more sophisticated model, improving relevance at the cost of slightly higher latency.

How It Works

Query → Retrieve (vector + graph) → Rerank → Final Results

Without a reranker, results are ranked by embedding similarity and graph scores. With a reranker, a cross-encoder or LLM evaluates each result against the query for more accurate scoring.

Setup

Add a reranker to your FlowRAG config:

import { LocalReranker } from '@flowrag/provider-local';

const rag = createFlowRAG({
  schema,
  ...createLocalStorage('./data'),
  reranker: new LocalReranker(),
});

The reranker is applied automatically to all query modes (hybrid, local, global, naive).

Implementations

Local (ONNX)

Runs entirely offline using a cross-encoder model:

import { LocalReranker } from '@flowrag/provider-local';

const reranker = new LocalReranker();
// Uses Xenova/ms-marco-MiniLM-L-6-v2 via ONNX

Pros: No API key, no network, fast, free
Cons: Less accurate than larger models

Gemini

LLM-based relevance scoring:

import { GeminiReranker } from '@flowrag/provider-gemini';

const reranker = new GeminiReranker();
// Requires GEMINI_API_KEY

Pros: High accuracy, understands nuance
Cons: Requires API key, higher latency, cost per query

AWS Bedrock

Amazon Rerank API:

import { BedrockReranker } from '@flowrag/provider-bedrock';

const reranker = new BedrockReranker();
// Uses amazon.rerank-v1:0

Pros: Managed service, good accuracy, AWS integration
Cons: Requires AWS credentials, cost per query

When to Use a Reranker

• Use it when result quality matters more than latency
• Skip it for simple lookups or when speed is critical
• Start without and add one if results aren't relevant enough