Grep or Embeddings? Agentic Search Over Company Documents — George He, LlamaIndex
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
Coding agents often search with grep, plain text matching, and skip , which index text by meaning. In this AI Engineer talk, George He, head of engineering at LlamaIndex, says that works for code because it is small, text-based and local, while company data is huge, messy, multimodal and permissioned, so the right choice depends on scale.
The proposal is a harness of five tools the agent chooses among: hybrid retrieval to orient, then listing, filtering and grep to dig in, and a fifth tool for reading complex documents through parsed text and page screenshots. The description says the talk demos an agent building a cash-flow table from 135 Alphabet filings. It is a vendor's argument: LlamaIndex makes the parsing tools involved. For the code side, Braintrust's talk Agentic Search vs Vector Search for Coding Agents takes up the same question.
- embedding — A list of numbers representing a piece of text's meaning, so that similar meanings end up numerically close and can be searched.
- Metadata — Data about a file, such as its name, size and timestamps, as distinct from its contents.
- Vector search — Finding content by similarity between embeddings.
- Harness engineering — Designing the full environment and feedback system around an agent.
Checking sign-in…
Loading comments…






