Grep or Embeddings? Agentic Search Over Company Documents — George He, LlamaIndex
LlamaIndex advocates hybrid grep-plus-embeddings over pure vector search for enterprise document agents
“Claude Code experimented with a vector database but ultimately abandoned it, and the reasons are not just performance, but also practical aspects of scaling and maintenance.”
George He of LlamaIndex (Series A) presents a practical framework for agentic document search, contrasting local grep-style file traversal against pre-indexed vector embeddings. He uses Claude Code's decision to abandon vector indexing—citing maintenance overhead from constantly changing code—as a real-world case study for when grep wins. The talk argues enterprise agents benefit from understanding which approach fits the data's mutability and structure, with a hybrid often being optimal.