Production RAG needs an operating model, not only a vector database

01

Retrieval is a governed information system

Retrieval-augmented generation becomes operationally important when people use it to find policy, prepare decisions, support customers, or act on enterprise knowledge. At that point the system inherits the ownership, permission, currency, and retention obligations of its sources.

A vector database can improve semantic access, but it does not determine which source is authoritative, who may retrieve it, when it becomes stale, or how a user should respond when evidence conflicts.

02

Preserve source authority and access boundaries

Every indexed collection should have a business owner, source system, access policy, refresh expectation, and retirement process. Retrieval must preserve the user and document boundary instead of creating a broad secondary copy with weaker controls.

Citations should point to the source material that supported the answer. They are part of the review experience, not decorative metadata.

• Authoritative source and named owner
• Document-level permission model
• Refresh, deletion, and retention behavior
• Citation and source-access path

03

Evaluate the complete answer path

Evaluation should test retrieval, generation, refusal, citation, permission, and escalation together. A fluent answer based on the wrong source is a system failure even when the language appears convincing.

Build representative questions from real workflows, including ambiguous requests, permission conflicts, missing evidence, outdated records, and cases where the system should defer to a person.

04

Operate model and knowledge change separately

Model, prompt, chunking, embedding, ranking, and source changes can each alter system behavior. Release records should identify what changed and compare the new behavior against the same evaluation baseline.

Production telemetry should show answer quality signals, unsupported responses, source coverage, retrieval latency, cost, user correction, and escalation. These signals connect the technical system to the workflow it is meant to improve.