Pharos
A local-first agentic RAG system that turns PDFs, scans, DOCX, PPTX and XLSX into a searchable team knowledge base with enterprise access control, exposed through both an HTTP question-answering API and an MCP interface an agent can reason over.
Pharos exposes one knowledge base through two doors that share a single tool contract and server-side identity checks so an agent cannot edit its own permissions. An HTTP API runs a closed retrieve-then-answer pipeline, backed by DeepSeek, for scripts and frontends that just want an answer with citations plus retrieval endpoints, while an MCP interface lets an agent such as Claude Code decide for itself when to retrieve, how to rewrite a query, and whether to search multiple times.
Because its embedded Qdrant vector store admits only one client and holds an exclusive lock, and its embedding model takes a minute or two to load, Pharos runs as a resident daemon that owns these heavy resources once, with MCP acting as a thin HTTP adapter that starts in milliseconds and every session sharing one already-warm backend, avoiding the lock contention and repeated model loads of a per-session process design. It combines multi-identity authentication, observability and systemd supervision, and installs as a single editable Python package; its code, CLI and commit history are in English, while its documentation, including a 12-part learning series, is written in Chinese.