← Back to projects

RAG Knowledge Engine

Retrieval pipeline that grounds LLM answers in a versioned, permission-aware knowledge base.

RAGVector DBLLM
Retrieval + answer demo

Problem

Generic RAG returns confident answers from stale or unauthorized documents. For an internal assistant, that is worse than no answer at all.

Approach

Retrieval is permission-aware and version-pinned: every chunk carries an ACL and a document version, and the ranker filters before the model ever sees a passage.

  • Hybrid search (dense + lexical) for recall on jargon and code.
  • Per-user ACL filtering pushed into the vector query.
  • Citations link back to the exact source version used.

Results

  • 40% fewer hallucinated citations versus the naive baseline.
  • Answers carry verifiable sources, so reviewers trust the output.