Obsidian Metadata
Is RAG Dead? No.
The viral post (“Why I No Longer Recommend RAG for Autonomous Coding Agents”, Nik Pash / Cline) argues against naive vector-DB retrieval for autonomous coding agents, not against RAG as a whole.
- RAG = Retrieval-Augmented Generation: using retrieval to give the model relevant context. The principle still holds: the LLM needs the right context. The question is how to retrieve.
- For code, naive vector similarity fails because code relationships are complex. Coding agents like Claude Code still retrieve, using agentic search (walking folders, reading imports, grepping) the way human developers do.
- Strategies range from keyword matching → embeddings → LLM relevance filtering → multi-hop agentic retrieval, and many production systems combine them.
- “RAG” has become a buzzword with no shared definition (any retrieval vs vector DBs only). Focus on the goal: getting the LLM the context it needs.
- Don’t follow categorical advice. Experiment with retrieval approaches and measure (see RAG Evaluation).
The Pash post itself: RAG is still fine for support bots, document Q&A, and memory. For high-quality coding agents, give the agent the tools engineers use and let it explore. “Retrieval is passive; agency is active.” Quote from Cursor’s Aman Sanger: “The hardest questions… require several hops. Vanilla retrieval only works for one hop.”
Hamel and Ben Clavié’s counterpoint: the 2023 approach (chunk → single vector → cosine) deserves criticism because single vectors lose information, but retrieval matters more than ever. LLMs are frozen at training time, and million-token windows don’t change the economics or fix context rot.