Ask your repo "where is X?" in plain English. Get the 3 files to open.
Runs locally · learns from your git history · ~80–100 ms per warm query · CLI + MCP
Real bug-report titles on a kubernetes clone; the file each fix changed ranks #1. About the demo.
curl -fsSL https://raw.githubusercontent.com/andreylukin/where-next/main/install.sh | shmacOS on Apple silicon and Linux with glibc 2.35+. The ~1.2 GB model downloads after the installer
asks. Read-first install, uninstall, Intel Mac and Windows status: docs/install.md.
Uninstall everything: wn uninstall.
cd your-repo
wn init # index the repo, learn from its git history
wn setup # connect Claude Code, Codex, Cursor: hints arrive automatically
wn ask "where are gitignore rules matched against paths"
wn bench # optional: replay past commits, see how it does hereOn a clone of ripgrep:
$ wn ask "where are gitignore rules matched against paths"
crates/ignore/src/gitignore.rs 0.50
crates/ignore/src/dir.rs 0.42
crates/ignore/src/overrides.rs 0.42
Scores rank the files; they are not probabilities. When nothing clears a calibrated threshold, wn
says "no confident hint". Query tips: docs/quickstart.md.
For agents, run wn setup: Claude Code, Codex and Cursor get the skill and hooks that add hints
to their context. See docs/skill.md; MCP (wn mcp) is there for other clients.
Grep needs the string you already know. Plain embedding search matches text that looks similar.
wn's model is fine-tuned on ~1.1M (task → files that actually changed) pairs, and a per-repo
adapter fitted on your commits in seconds learns your repository.
hit@3 = a file the real fix changed is in the top 3. ContextBench, official 500-task subset.
¹ BM25 is on all 1,136 tasks. On the 994 tasks from repositories held out from training: untrained
.46, wn .76, with adapter .80. Protocols and more models: benchmarks ·
FAQ.
On 102 real closed issues, the title alone put a fixed file in the top 3 49% of the time vs 31% for grepping its identifiers; with full bodies grep is ahead (issue titles).
- No measured agent savings. Four controlled trials found no lower cost or higher success for a capable agent; even handing it the files the real fix touched barely helped (agent trials). Use it as navigation, not a cost-saver.
- Exact names and strings: use rg.wnranks by meaning.
- Vague follow-ups like "now the other one" get "no confident hint" rather than a guess. Ask full questions.
- Early. v0.1.0. Windows and Intel Macs are not supported yet.
Full list: docs/limitations.md.
Docs · FAQ · Troubleshooting · Benchmarks · Changelog · Contributing
Code: Apache-2.0 (LICENSE). The default model,
lukandrey/where-next-gemma-xl1, is
distributed separately; it is fine-tuned from google/embeddinggemma-300m and is subject to the
Gemma Terms of Use. See NOTICE and
privacy and licensing.
No telemetry. Your code never leaves your machine. wn report shares anonymous usage stats only
after you read and confirm them (docs/stats.md).