oad/search-doc

Search documents/ semantically only for product-capability or repository-policy questions. The codebase, tests, CLI help, and UI are authoritative for implementation and operation; use grep, glob, source reads, and git history for implementation and historical rationale.

This wraps the experiment at experiments/turbovec_docs/ (turbovec index + local sentence-transformers embeddings). Reuse the index when current; rebuild it only after the retained Markdown corpus changes.

1. Check the index exists

ls experiments/turbovec_docs/index/docs.tvim 2>/dev/null

• If missing, or if documents/ has changed since the last build, rebuild it:

uv run --group experiment python experiments/turbovec_docs/build_index.py

The corpus is intentionally small: the feature catalogue and repository instructions. Rebuild whenever it might be stale rather than treating old search results as authoritative.

2. Search

uv run --group experiment python experiments/turbovec_docs/search.py "<query>" -k 5

• Use natural-language questions about shipped features or repository documentation policy.
• -k controls result count (default 5); lower it when one clear result is expected.
• Open the returned file and line range before quoting or acting on a result.

3. Use source search instead when

• You need implementation behavior, API routes, schemas, configuration keys, CLI arguments, UI details, tests, or error strings.
• You are looking for an exact identifier, function/variable name, or literal text.
• The semantic result is low-confidence or the retained corpus does not cover the question.

Notes

• This is local and offline after the one-time model download.
• experiments/turbovec_docs/README.md describes the experimental index and its limitations.
• The experiment uv dependency group is dev/local-only; never assume it exists outside this workspace.