General RAG System
An enterprise, multi-tenant Agentic RAG knowledge-base system with a Vue.js frontend, a Spring Boot backend and a FastAPI LLM service, where the model chooses its own retrieval tools across several rounds and exposes knowledge bases to agents over MCP.
General RAG System lets a language model decide, turn by turn, which of six retrieval tools to call and when to stop retrieving, rather than always performing one fixed retrieval step; a gateway first decides whether retrieval is needed at all, falling through to a plain language-model response when it is not, and the system's only retrieval path is this agentic, multi-round one. Retrieved documents are deduplicated, and adjoining passages can be read together for more context, up to ten rounds per conversation by default.
It exposes an authorized subset of a knowledge base's capabilities to external AI agents over MCP using the Streamable HTTP protocol, authenticated with per-key access tokens that are shown once and stored only as a hash, and audits every tool call asynchronously. Multi-tenant workspaces isolate knowledge bases with owned, workspace-shared, invited and public visibility levels and role-based membership; uploaded documents are deduplicated by a checksum, parsed and chunked by file type, and stored as vectors in Milvus, with authentication through JWT tokens checked against Redis session state.
Its frontend is Vue 3 with Ant Design Vue, its backend is Spring Boot with Java 11, its LLM service is FastAPI with LangChain, and it uses MySQL, Redis, RabbitMQ and MinIO; it is licensed under Apache-2.0.