v0.9.0

Changelog

Features

• Ollama auto-start and auto-pull - When the embedding backend is unavailable at startup, quant automatically starts Ollama (ollama serve) if it is installed locally and the embed URL points to localhost, then pulls the configured model if it is missing. Both recovery steps happen before the MCP server opens, so clients connect to a fully operational server without manual setup.
• Keyword-only degraded mode - If embedding recovery fails (Ollama not installed, remote URL unreachable, or network error), quant starts without an embedding backend. The MCP server is fully operational; search falls back to FTS5 keyword results and index_status reports the embedding status and the fix needed.
• OpenAI-compatible embedding provider - A new openai provider supports any API that follows the OpenAI embeddings contract. Select it via --embed-provider openai (or QUANT_EMBED_PROVIDER=openai); auto-detected when embed_url contains openai.com.
• Embedding API key support - New --embed-api-key flag and QUANT_EMBED_API_KEY environment variable pass a Bearer token to the embedding backend, enabling authenticated providers such as OpenAI.
• drill_down MCP tool - Explores a topic by finding diverse chunks related to a seed chunk, spreading results across different source files rather than staying within one document.
• summarize_matches MCP tool - Returns a high-level overview of which documents match a query and what they contain, without returning individual chunks.
• list_collections and delete_collection MCP tools - List named collections with document and chunk counts; delete all content in a named collection.
• Embedding status in index_status - The index_status tool now reports whether the embedding backend is available or the server is running in keyword-only mode.
• Chunk summarization - New --summarizer flag enables LLM-powered summarization of indexed chunks at ingest time. The summarization model is configurable via --summarizer-model.
• Cross-encoder reranking - New --reranker cross-encoder flag adds a cross-encoder reranking pass after hybrid retrieval. The reranker model is configurable via --reranker-model.

Improvements

• Actionable Ollama error messages - Connection errors and 404 model-not-found responses now carry human-readable messages with the exact ollama serve / ollama pull commands to run.
• Ollama process isolation - When quant starts Ollama automatically, the child process is placed in its own process group (Setpgid: true on Linux/macOS) so Ctrl+C in the terminal does not kill Ollama along with quant.