RAG-Intel (open source)

Transparent local retrieval engine where every ranking decision is visible

Overview

An MIT-licensed retrieval engine extracted from a larger document-intelligence system: bounded chunking, BM25 with optional SentenceTransformers + FAISS, explicit reciprocal-rank fusion, SQLite state, and optional cross-encoder reranking. The ranking policy is explicit TOML rather than hidden defaults, every result carries its matching evidence chunks, and the lexical path runs fully offline. It is a reference engine, not a framework and not a service; no performance claims are made.

Key Highlights

• Bounded chunking, with overlap counted inside the requested size
• Retrieval policy is explicit TOML — no silent ranking defaults
• BM25 and optional dense retrieval fused by an explicit strategy
• Results carry document identity, scores and the matching chunks
• Offline lexical path: no model download, no network access

Tech Stack

Links