2026-09-21 · America/Los_Angeles · 社区动态 · #16

A year of reading AST-aware code RAG papers, and the thing I built

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: -…

热度 45.6 / 100；排名与评分保留该期记录。