Sourcelens

[English](README.md) | [中文](README.zh-CN.md)

**Harness-based Agentic RAG** — no embeddings, no vector DB, no pre-indexing

**SourceLens** is Agentic RAG built on an AI coding agent harness — the same kind of harness behind tools like Cursor, Claude Code, or Codex, not those products themselves — running inside a sandboxed environment. Instead of embedding your files into a vector index ahead of time, SourceLens hands them directly to the agent harness, which reads, searches, and reasons over the file system on demand — turning any pile of documents or code into something you can just ask questions of.


Instead of vector embeddings or keyword indexes, SourceLens uses AI coding agents running in a sandbox to directly read, navigate, and reason over the file system. This means the retrieval understands code structure, cross-file relationships, and semantic intent — not just surface-level text matching.

Background

Our first attempts at RAG used graphical workflow tools like Dify and n8n. They asked a lot of the people building on them, and the real difficulty was always upfront: splitting documents and embedding them before they ever reached a vector store. That prep work took real effort to get right, and even after all of it, recall accuracy stayed disappointing — answers would come back incomplete, sometimes missing the point that was in the document all along.

Around the same time, we noticed something different from using Cursor for development: it did no pre-training or pre-indexing at all, yet it was consistently accurate at reasoning over a codebase. That raised an obvious question — why not use the same approach for RAG?

That's the idea behind Sourc