# I made a 5-33% cheaper code search tool for AI agents (sweet-search) — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/e2a9d52240b0d175

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-11-02 14:31 ingested 2026-11-03 02:42

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1xlyqvf/i_made_a_533_cheaper_code_search_tool_for_ai/)

## retained text



```
Hi guys!

I've spent most of the past year fighting windmills. The result is [sweet-search](https://github.com/mrsladoje/sweet-search) \- local code search for AI agents (Claude Code, Codex, opencode, ...).

*What does sweet-search do?*

sweet-search makes an index of your repo (and then automatically maintains it) and gives the agent 6 code retrieval tools.

Benchmark results (200 held-out SWE-rebench tasks, 5 combos harness + model):

* **5–33% cheaper runs** (more on [GitHub](https://github.com/mrsladoje/sweet-search#-benchmarks))
* **same no. of solved tasks** (in total): 481 vs 481 (of 1000)
* Opus 5.5 & Sol 6.1 (not some weak models)

*The technical part:*

* Rust core, native binary for your OS-CPU combo
* custom MaxSim and ANN (generally Pareto better than competitors like FAISS by Meta)
* *ss-grep* (custom grep tool - 2.46× end-to-end faster than *ripgrep*, with custom ranking of hits)
* custom attention layer for embedding and late-interaction models (for GPU acceleration for indexing)
* 100% local

GitHub: [https://github.com/mrsladoje/sweet-search](https://github.com/mrsladoje/sweet-search)

Wanna hear what you guys think :)
```



document fetched 2026-11-03 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1xlyqvf/i_made_a_533_cheaper_code_search_tool_for_ai
