Function-Level Chunking Significantly Improves LLM Code Retrieval Accuracy

A developer building a retrieval-augmented generation (RAG) system over a codebase found that splitting source code by fixed token counts produced fragmented, near-useless results. The core issue was that standard line-based chunking broke functions mid-body, leaving embeddings that represented incomplete code with no signature or context. Switching to function-aware chunking — one complete function per chunk, including its signature and doc comment — dramatically improved the quality of answers from the language model. The approach uses syntax tree parsers like ts-morph for TypeScript or AST-based tools for Solidity to extract whole function nodes rather than raw text slices. Running the embedding pipeline locally via Ollama ensured that private codebases never left the developer's machine while still enabling accurate, name-based semantic queries.