A document assistant that can cite its sources.

The need and its boundaries

Imagine a support team working with product procedures. The assistant should retrieve relevant passages, draft a response and let the person verify its sources. It does not modify accounts or carry out operations for a customer.

The initial scope uses a defined, maintained document set. Personal documents, unapproved sources and content outside the user’s permissions are excluded from retrieval.

A five-step workflow

Retrieval-augmented generation, or RAG, combines source retrieval with response generation. These stages have different responsibilities.

• Authenticate the person and determine their permissions.
• Retrieve relevant passages from the sources they are allowed to access.
• Send a limited selection of passages and their identifiers to the model.
• Produce an answer with verifiable references, or flag insufficient information.
• Show the answer and sources to the user without automatically acting on business systems.

Controls belong in the software

The backend enforces permissions before documents are accessed. An instruction to the model does not replace access control. Retrieved content is treated as data, not as instructions to execute.

Document changes and deletion must propagate to the index. Displayed references should point to relevant versions that the person asking the question can access.

Test more than the easy answers

The evaluation set includes questions with known answers, but also contradictions, outdated documents, out-of-scope questions and information the requester is not allowed to access.

Retrieval quality, faithfulness to sources and the ability to abstain are assessed separately. A fluent answer is not necessarily correct. The findings inform whether the scope can be expanded.

Plan for changes and incidents

Operations include usage limits, error monitoring, a way to report incorrect answers and a fallback to conventional search. Logs should not retain sensitive content without a clear need.

Changes to the document set, retrieval system and model are evaluated before release. This architecture is a starting point: final decisions depend on the data, risks and product that hosts the assistant.