Agentic RAG for local and self-hosted document search: hybrid retrieval, reranking and multimodal RAG on embedded LanceDB, with Docling parsing and an MCP server

Recommends the right Claude model for a task — Haiku, Sonnet, Opus or Fable — at the start of work and reviews the choice afterward to suggest a lighter or stronger option, running locally with no API calls or token cost.