Reranking

Reranking is a crucial step in the recall pipeline. After retrieving a set of potentially relevant documents (e.g. via keyword search or basic vector search), a reranker uses a more powerful semantic model to re-order those documents by their true relevance to the query.

Core Package

@tekmemo/rerank provides the standard Reranker contract and a deterministic fallback implementation.

Install

npm install @tekmemo/rerank

API Reference

Method	Purpose
reranker.rerank(input)	Scores and sorts documents by relevance.
createDeterministicFallbackReranker()	Creates a local reranker that sorts based on keyword presence (useful as a fallback).

---

VoyageAI Adapter

@tekmemo/rerank-voyage provides a production-ready implementation of the reranker contract using VoyageAI's specialized rerank models.

Install

npm install @tekmemo/rerank-voyage

Example usage

import { createVoyageReranker } from "@tekmemo/rerank-voyage";

// Create the reranker instance
const reranker = createVoyageReranker({
  apiKey: process.env.VOYAGE_API_KEY,
  model: "voyage-rerank-2"
});

// Re-order recall results for maximum relevance
const results = await reranker.rerank({
  query: "How do I handle sync conflicts?",
  documents: [
    "Conflict resolution policy: keep-cloud...",
    "Sync push sends local events to the cloud...",
    "Memory records should be small and explicit."
  ],
  topK: 1
});

console.log(`Top match: ${results[0].document}`);

Use Cases

• Quality Improvement: Use a cheaper, faster search (like keyword or small vector) for the first pass, then use a reranker to ensure the top results are the most relevant.
• Context Window Optimization: Ensure that only the absolute best information is injected into your agent's context window.
• Hybrid Recall: Combine local keyword results and cloud vector results, then rerank the union for a unified set of context hits.