RAG for Developers — Built for Code, Not Just Text (Review Requested)

Code-Aware RAG Tool — Looking for Developer Feedback

We’ve been building a code-first RAG tool that actually understands how codebases work, not just how text looks in embeddings. The goal is simple: when you ask a question, you get the right functions, related calls, and supporting code, not random nearby snippets.

What’s Inside

AST-based code chunking with Tree-sitter (Python, JavaScript, TypeScript)
Functions, classes, imports, calls, docstrings — all extracted explicitly

A clean async ingestion pipeline with strict tool → agent → storage boundaries

Semantic vector search as the starting point, not the end

In-memory dependency graph expansion:

Built lazily from chunk metadata

No persistence, no globals, no backend shortcuts

Stable qualified IDs (file::entity)

Context expansion via BFS over calls and imports to pull in code that’s actually connected

Backend-agnostic vector store layer, so storage can change without rewriting logic

Why We Think This Is Useful

You get related code paths, not just similar text

Context stays small, relevant, and debuggable

The architecture avoids hidden state and scaling surprises

What We’d Love Feedback On

If you’ve worked with large repos or built RAG systems before, we’d really value your thoughts on:

The “graph as derived state” design

Chunk metadata choices (calls, imports, QIDs)

Retrieval + expansion flow

Any edge cases you think would show up in real production codebases

Even quick reactions or gut checks are welcome.