Ran an Opus agent over a set of documents twice, once with plain Grep, Glob and Read and once with Semlith.
With grep it found 63% of the passages, with Semlith 98%. The part that surprised me is that 41 of its 127 grep quotes weren't actually in the document it said they came from. With Semlith that was 0 of 126.
On a single code repo it made no difference, 92% either way. Has anyone else caught their agent quoting stuff that isn't there?
https://github.com/semlith/semlith
GitHubGitHub - semlith/semlith: A fast local vector store and code graph for AI agents — index your files once, keep it current as you save, and answer questions across all of it in milliseconds without leaving the machine.A fast local vector store and code graph for AI agents — index your files once, keep it current as you save, and answer questions across all of it in milliseconds without leaving the machine. - sem...
commit 865b704