Enterprise Knowledge Fusion RAG

A modular, multi-source retrieval-augmented generation platform that ingests PDFs, crawled websites, CSVs, and read-only SQL databases, then answers questions with grounded, inspectable citations.

Key Capabilities

• Three retrieval representations rebuilt on every ingest: vector search, sentence-window context, and a durable local BM25 lexical index for exact identifiers and terminology.
• Reciprocal rank fusion combines multiple retrievers into one deterministic ranked list without mixing incomparable raw scores across strategies.
• Opt-in cross-encoder re-ranking scores a larger fused candidate pool and applies a per-source diversity pass before returning the final results.
• Optional knowledge graph retrieval with deterministic two-hop traversal and edge-level citations spanning every supporting source.
• Production persistence fails closed unless PostgreSQL, Azure Blob Storage, and Pinecone are fully configured, with atomic generation publishing so readers never see a partial write.
• Enterprise hardening: role-gated organizations, a hash-chained audit log, per-workspace retention policies, source security classifications, rate limiting, and prompt-injection scanning.

Tech Stack