hybrid-rag-router

Intelligent hybrid RAG router for vector, SQL, document, and memory stores.

What changed in 0.2.0

This version adds store weighting, route confidence scoring, and tracing hooks so developers can make routing decisions more transparent and tunable.

Why

Most RAG systems need more than one retrieval method. Some questions are better answered from SQL, some from policy documents, and some from memory or semantic search.

Install

npm install hybrid-rag-router

Usage

import { HybridRouter } from 'hybrid-rag-router';

const router = new HybridRouter([
  {
    name: 'vector-main',
    type: 'vector',
    weight: 1.1,
    query: async (query) => [{ text: 'Vector chunk', source: 'vector-store', score: 0.9 }]
  },
  {
    name: 'sql-main',
    type: 'sql',
    weight: 1.4,
    query: async (query) => [{ text: 'Structured row result', source: 'postgres', score: 0.8 }]
  }
], {
  onTrace: (event) => console.log(event)
});

const route = router.route('What is the total approved leaves this quarter?');
console.log(route.intent, route.confidence);

const context = await router.getContext('What is the total approved leaves this quarter?');
console.log(context);

Features

• Query intent classification
• Store routing by type
• Store weighting for ranking
• Route confidence scoring
• Trace hooks for debugging
• Merge and deduplicate chunks
• TypeScript types included

Supported intents

• structured -> prefers SQL stores.
• policy -> prefers document and vector stores.
• personal -> prefers memory stores.
• factual -> prefers vector and document stores.
• hybrid -> falls back to all stores.

Next ideas

• LLM-based routing
• Reranking plugins
• Evaluation helpers
• Routing analytics