grepticon/

Benchmark

How a filesystem plus grep compares to naive RAG on answer accuracy — the headline result and where to read the full write-up.

Grepticon replaces an embedding index with a filesystem and grep. The obvious question is whether that retrieves as well. On answer accuracy, it retrieves better.

The headline

Across an agentic evaluation on a document corpus, giving the agent ls / find / cat / grep over the workspace beat a naive top-k RAG baseline on answer accuracy:

  • +16.7 points on the hosted stack.
  • +21.9 points in the in-lab run.

The agent compensates for vocabulary gaps by searching stems and synonyms — the way a person greps a codebase — and reasons over whole files instead of context-window-sized chunks, which is where naive chunked RAG tends to lose the thread.

Read the full write-up

The methodology, the systems compared, the corpus, and the complete numbers live in one canonical place — the benchmark write-up on the marketing site:

The agent filesystem vs. RAG →

That write-up is the single source of truth for the figures; this page just summarizes the result and points you there.

On this page