Writing
Blog
Benchmarks, engineering notes, and the thinking behind giving agents a filesystem instead of a vector database.
Benchmark · July 23, 2026
A filesystem beats naive RAG on retrieval accuracy
We put a grep-over-filesystem agent against top-k and agentic RAG on an adversarial 18,340-file corpus. On the production stack it answered 97.9% correctly — 16.7 points over naive RAG. Method, numbers, and honest caveats.
6 min readRead the write-up →
+16.7points vs naive RAG
More writing soon.