v0.3.0

What's New

HTTP Serve Command

New vault-recommender serve subcommand starts an HTTP server on port 7532 (configurable) that loads the index once and serves fast queries. Designed for Claude Code hook integration where sub-second latency is required.

Endpoints:

• GET /health — readiness check
• GET /recommend?topic=X&top_k=3 — semantic search
• GET /recommend?note=X&exclude_linked=true — note-based recommendations
• POST /reload — hot-reload index from disk

MCP Server Improvements

• Eager loading via FastMCP lifespan — index and link graph load at startup instead of on first tool call
• Uses ctx.lifespan_context for clean state management

Other Changes

• create_recommender() factory shared across CLI, HTTP server, and MCP server
• Removed unused scikit-learn dependency (cosine similarity uses numpy dot product)
• 39 tests, all passing

v0.2.0

What's new

• --rebuild flag: recommend --rebuild checks if any vault file is newer than the index and auto-rebuilds if stale. Skips silently if fresh.
• reload_index MCP tool: Force-reload the index from disk without restarting the MCP server.
• Improved text preparation: Labeled sections (Topic/Title/Tags/Related/Body) and wiki-links included in embeddings for better semantic matching.
• Tuned scoring: link_boost_1hop increased from 0.15 to 0.50 based on autoloop optimization.
• Index validation: VaultIndex now validates that entries and embeddings have matching lengths on load.
• Updated README: PyPI/CI badges, --rebuild docs, full tool listing.

v0.1.1

Fix slow CLI startup — defer heavy imports so --help responds instantly instead of waiting ~2.5s for torch to load.

v0.1.0

Initial release