Karpathy's knowledge base matches our Grep-is-All-You-Need paper
a day ago
- #knowledge grounding
- #LLM agents
- #retrieval-augmented generation
- The paper proposes Knowledge Search, a two-layer retrieval system using grep and cat, as an alternative to the standard RAG stack for domain-specific knowledge grounding.
- It claims 100% retrieval accuracy with sub-10ms latency, zero preprocessing, and no additional memory or infrastructure dependencies, based on deployment in 20 LLM agents across three knowledge domains.
- Key advantages over vector RAG include no need for embedding model selection, chunking, vector databases, or approximate nearest neighbor search, reducing complexity and latency.
- Limitations include unsuitability for open-domain queries, semantic similarity searches, cross-lingual retrieval, and very large corpora beyond ~1GB.