[dominikpeter]

Closes #12.

What changes

• New preset gemma (embeddinggemma through a local ollama, loopback only). A bare --embed now means gemma; the other presets are written --embed=minilm|qwen3|potion|openai|cohere.
• README: recommends ollama pull embeddinggemma && duplicatecode scan . --embed; MiniLM is no longer described as best.
• Benchmark tooling for Recommend a default light embedding model: quality and speed across codebases #12: eval/make_codebase_task.py, eval/embed_codebases.py (new ollama:<model> backend), eval/analyze_codebases.py, eval/static_rank/.
• Version 0.3.2 (merging publishes it).

Why gemma

Re-implementation retrieval over 8 codebases, 240 functions (details in #12): R@1 0.942 vs 0.892 for the static detector alone, and still 0.942 with function names masked. MiniLM-L6 0.854, bge-m3 0.912, Qwen3-0.6B 0.917. About 5 ms per text on a GPU.

Tested

• cargo test --workspace --locked (52 pass), cargo fmt --check.
• eval-groups on the holdout, default profile: static 0.7607 -> 0.9071 with --embed (gemma). Per set: fn-python 0.706 -> 0.801, fn-typescript 0.728 -> 0.920, javascript 0.497 -> 0.788, python 0.660 -> 0.948, react 0.980 -> 0.987, sql 0.993 -> 0.9995.

Not covered

Default is unchanged without --embed. Thresholds with embeddings were set for MiniLM; false-positive rate of gemma on a real scan is not measured. The 11 HF-only models were not run.

🤖 Generated with Claude Code