The second public alpha of docmd-search. This release brings the TUI, the browser-ready index format, and the engine-task integration. APIs and behaviour may still change before 0.1.0.

What’s in This Release

TUI for index management and search

A first-class terminal interface for the indexer, model manager, and search runner. First run walks you through a setup wizard; subsequent runs land in a search TUI. Settings (model, cache dir, project root) are reachable from a settings panel inside the TUI.

Embedding model support

Loads models from the @huggingface/transformers registry. Default is Xenova/all-MiniLM-L6-v2 (English-only, ~23 MB). Models are downloaded on first use and cached in ~/.cache/huggingface.

Browser-ready index

Index is written as a single manifest.json plus N batch files of vector records. The search plugin’s client code reads the manifest, downloads only the batch it needs, and runs cosine-similarity ranking in the browser via the same model. No server round-trip per query. Batches are Int8-quantised by default.

Engines

Uses the same engine task protocol as the rest of the docmd ecosystem: search:chunk for splitting pages into overlapping windows, search:quantize for Int8 per-vector quantisation, search:cosine for batch scoring. Rust engine is used when present, then JS engine, then a built-in inline implementation. No engine is required.

Known Limitations

• English-only default model
• No incremental updates — full rebuild required after any content change
• Browser memory usage roughly 50–100 MB during a query
• First load is slower as the embedding model is fetched and cached
• Dist-tag is alpha until 0.1.0 ships

What’s Next

• 0.1.0 (stable) will add the default condition to exports, adopt the docmd namespace, and switch the publish workflow off --tag alpha so the dist-tag becomes latest.