Chonks：基于AST与混合检索的多语言代码RAG工具

Chonks indexes a large codebase into one SQLite file and serves it to a coding agent over MCP. Tree-sitter AST chunks, hybrid semantic plus BM25, a typed cross-reference graph. No LLM in the retrieval path. Apache-2.0, free, no telemetry. Built for multi-million line codebases.

Ten languages get full symbol and graph support, including C++, C#, HLSL and Python, so it works where a call path crosses languages.

Acc@5 64/100 on held-out Loc-Bench, in one call. The ids were picked before tuning and are published in the repo.

The reading list, and what I got from every single one:

• [cAST](https://aclanthology.org/2025.findings-emnlp.430/), Findings of EMNLP 2025. Recursive AST-node splitting with sibling merging. The core chunking algorithm.
• [LARGER](https://arxiv.org/abs/2605.16352), 2026. Lexical anchors, then confidence-filtered expansion. Became my graph seed gate.
• [Reformulate, Retrieve, Localize](https://arxiv.org/abs/2512.07022), 2026. Pull identifiers and stack-trace terms out of bug report text before retrieval. I do it rule-based, with no LLM call.
• [From Fragments to Paths](https://arxiv.org/abs/2606.22906), 2026. Budget-constrained traversal over multi-relational repository structure. The same shape as my graph expansion.
• [AST-derived vs LLM-extracted graphs](https://arxiv.org/abs/2601.08773), 2026. The head-to-head behind my choice to keep the LLM out of index time.
• [Practical Code RAG at Scale](https://arxiv.org/abs/2510.20609), 2025. The counterpoint. Line chunking matches syntax-aware splitting on completion benchmarks, and BM25 beats dense embeddings for code-to-code queries.
• [LocAgent](https://arxiv.org/abs/2503.09089), 2025. Released Loc-Bench, which I use as the held-out set.
• [GraphRAG](https://arxiv.org/abs/2404.16130), 2024.

Most of these I found while building. A lot of the time it turns out, what I was building was converging to the lit.

https://github.com/mgonzalez01/Chonks