Search is the foundation. Summarization and MCP access are what turn a unified index into something you can actually use.
Every query can combine full-text relevance (OpenSearch BM25, tuned against a real evaluation set rather than left on defaults) with optional vector similarity search. The two are merged by score, not just rank position, so a strong exact-term match and a strong semantic match both get weighed on how confident they actually are.
Vector search is optional, not a requirement. Running an embedding model is real resource cost, so the text-only path works on its own - including on hardware as modest as a Raspberry Pi. Turn on vector search when you want it, using a local model (Ollama) or a hosted provider you configure yourself.
Point a search at a configured model and get a written summary of the top results instead of a list to read through yourself - useful when you want the answer, not another page of links. You choose the model and the provider; nothing is summarized unless you've configured it to be.
Yew includes an MCP server so AI tools can search and read your indexed content directly, instead of you copy-pasting context into a chat window. It's early and still evolving - see the dedicated MCP page for exactly what works today versus what's still coming.
A planned assistant that can search, read, and reason over multiple turns - grounded in your own content with citations back to the source, not a generic chatbot. This doesn't exist yet; it's on the roadmap and will be built as a consumer of the same MCP server described above, not a second, separate retrieval system.