2279389415
- Scope package to @tobi/qmd, version 0.9.0 - Add files whitelist, publishConfig, repo metadata - Add CI workflow (bun tests on ubuntu + macos, bun latest + 1.1.0) - Add publish workflow (triggers on v* tags, publishes to npm) - Add release script for version bumping + changelog generation - Add LICENSE (MIT) and initial CHANGELOG.md - Update install instructions to use @tobi/qmd
1.6 KiB
1.6 KiB
Changelog
All notable changes to QMD will be documented in this file.
0.9.0 - 2026-02-15
Initial public release.
Features
- Hybrid search pipeline — BM25 full-text + vector similarity + LLM reranking with Reciprocal Rank Fusion
- Smart chunking — scored markdown break points keep sections, paragraphs, and code blocks intact (~900 tokens/chunk, 15% overlap)
- Query expansion — fine-tuned Qwen3 1.7B model generates search variations for better recall
- Cross-encoder reranking — Qwen3-Reranker scores candidates with position-aware blending
- Vector embeddings — EmbeddingGemma 300M via node-llama-cpp, all on-device
- MCP server — stdio and HTTP transports for Claude Desktop, Claude Code, and any MCP client
- Collection management — index multiple directories with glob patterns
- Context annotations — add descriptions to collections and paths for richer search
- Document IDs — 6-char content hash for stable references across re-indexes
- Multi-get — retrieve multiple documents by glob pattern, comma list, or docids
- Multiple output formats — JSON, CSV, Markdown, XML, files list
- Claude Code plugin — inline status checks and MCP integration
Fixes
- Handle dense content (code) that tokenizes beyond expected chunk size
- Proper cleanup of Metal GPU resources
- SQLite-vec readiness verification after extension load
- Reactivate deactivated documents on re-index
- BM25 score normalization with Math.abs
- Bun UTF-8 path corruption workaround