Hi everyone :) After almost a year of fighting windmills, I am proud to say I built sweet-search, a local code search tool for Claude Code, Codex and Opencode.
It delivers 5-33% cheaper agentic runs with Sol 6.1 and Opus 5.5 with Claude Code, Codex and Opencode, with pooled solved tasks equal vs native (both 481 of 1,000 tasks pooled on 5 model*harness combos). Measured on 200 held-out SWE-rebench tasks. Detailed bench results are in the README.
It consists of 6 tools: hybrid search with late interaction reranking, my version of grep, and others. It also has GPU accelerated indexing (for embedding and late interaction models) and low footprint incremental indexing.
Some highlights: 1. The ANN kernel is custom and very competitive with FAISS and USearch. 2. The custom MaxSim kernel is 1.4–7.3× faster than NumPy and PyTorch (reranking on CPU). 3. My hybrid-search tool is SOTA on M2CRB and zero-shot SOTA on CoSQA. 4. My custom grep tool is 2.46x faster e2e than ripgrep and ranks hits in a specific way (inspired by Sainte-Laguë).
Here is the promo video I built with fframes and Opus 5.5: https://www.youtube.com/watch?v=lJzgXRQprs4 Would love to hear your opinions, Marko :)
Comments URL: [https://news.ycombinator.com/item?id=50326635](https://news.ycombinator.com/item?id=50326635)
Points: 2