Answer from enterprise knowledge without crossing access boundaries

Employees need answers drawn from scattered documents. The organization cannot let an assistant expose restricted records or confidently combine contradictory policies. Retrieval relevance, current policy versions, document permissions and answer support are separate problems. A citation can point to a real document and still fail to support the answer it sits beside.

We build an answer system that checks access first and ties every answer to exact source passages.

• The system authenticates the caller and applies authorization before any content enters model context, using source permission metadata and user and group identities.
• It retrieves relevant passages from approved documents, resolves which version is current, and generates an answer linked to the exact passages it uses.
• Conflicting or missing evidence is shown as a reviewable gap. If retrieval finds no adequate evidence, the system declines to answer and routes the question to the knowledge owner.
• Access checks apply to every passage, citation, saved answer and log. If permissions are out of date, those sources are blocked until they are confirmed, and cached answers are cleared when access is revoked.
• We decide where embeddings, indexes, answers and logs reside. Reusing a private model does not by itself preserve source permissions.

• Answers with source excerpts. Each answer is tied to the exact passages that support it, from the current version of each document.
• Explicit conflict and gap states. When sources disagree or do not cover the question, the system says so instead of guessing.
• An audit trail. The permission decision and source versions behind each answer are recorded.

• Knowledge owners approve document precedence and reference answers.
• Access administrators verify the identity mapping and source permissions.
• Users can flag unsupported answers, and sensitive decisions still follow the underlying approval process.

Use ordinary enterprise search when finding the right document is enough. If policies are contradictory or unmaintained, improve document ownership before adding generated answers.

• Intent-aware retrieval for enterprise RAG: lessons and the data it produces
• Approved and generated answers: the data a two-tier help assistant can give you