A Hacker News thread asked developers how they implement Retrieval-Augmented Generation locally with minimal dependencies for code and documents. The community offered diverse solutions, from traditional search to vector databases, reflecting a broader push for private, offline AI. The discussion revealed a split between simple, proven tools and more complex, AI-native systems.

Developers split on the best approach for code versus documents. Many argued that BM25 and trigram search outperform vector embeddings for code, citing speed and accuracy. For documents, solutions like SQLite FTS5 and hybrid Postgres setups were praised. Some preferred lightweight tools like Ollama and Ripgrep, while others leaned on dedicated platforms like AnythingLLM.

The core debate centered on complexity versus utility. A key question emerged: why use RAG at all when direct tooling often yields better results? While vector search is powerful for text, its overhead and noise make it ill-suited for code without heavy tuning. The conversation highlighted a preference for pragmatic, fast solutions over pure AI hype.