Role Aware RAG Platform
A role-aware retrieval-augmented generation platform for enterprise knowledge bases that filters search results by the requesting user's role and document permissions.
This platform applies a role-based permission filter at the retrieval entry point, so a user only retrieves documents they are allowed to see, supporting admin, manager and staff roles. It is aimed at internal knowledge search, design-document question answering and other business systems that need access control, and combines vector search, BM25 and reciprocal-rank-fusion hybrid search with an optional reranking stage, selectable at runtime.
A custom Excel parser targets Japanese-style specification documents that use Excel as a layout tool rather than a data table, preserving sheet structure, merged cells, key-value regions and table regions that a direct text conversion would otherwise lose, alongside a unified loader for Excel and PDF formats. Answers return the source text, matched chunks, token counts and latency for traceability.
The backend is FastAPI with SQLAlchemy and Alembic, retrieval runs on Qdrant with local BM25, reciprocal rank fusion and reranking, and the default AI stack points an OpenAI-compatible client at a local Ollama server running BGE-M3 embeddings and a Qwen model; an offline evaluation on a public benchmark subset reported 90.9 percent accuracy over 132 questions.