Tag
A new open-source tool called Writ uses a hybrid retrieval pipeline with BM25, ONNX vectors, and Neo4j graph traversals to provide context rules for AI coding agents, reducing token bloat by 726x and enforcing plan approval via bash hooks.
The authors detail their experience building a code indexing system, concluding that graph-based retrieval with LLM-generated semantics outperforms vector embeddings and pure AST parsing. They open-sourced the system, Bytebell, which uses Neo4j to store semantic context for efficient and precise code retrieval.