Enterprise RAG: answers with sources, permissions and current data

The useful question: what evidence supports the answer?

Retrieval-augmented generation, or RAG, connects a question to relevant content before generating an answer. Microsoft describes this pattern as a way to ground answers in proprietary content [1]. It helps query policies, procedures and documented incidents. The goal is to reduce manual searches without turning a plausible answer into an automatic decision.

For this October 2026 edition, the priority is to move from a search demonstration to a service with identity, updates and evaluation. Adding documents or changing the model does not, by itself, fix a permissions, versioning or evidence problem.

Apply permissions before retrieval

The authenticated identity and its groups determine eligible documents. The service applies restrictions before selecting passages and building the model context. Filtering only on-screen citations is insufficient: restricted content would already have reached the model. Microsoft documents query-time controls using permission metadata; some implementations are in preview and require checking their limitations [2].

• Keep the document identifier, owner, version, classification and access rules alongside its passages.
• Resolve identity on the server; do not trust a group or tenant freely supplied by the browser.
• Revalidate permissions when opening a source and isolate caches by identity and access context.
• Reject retrieval when permissions are missing or their validity cannot be verified.

An index also needs a retirement policy

Storing the ingestion date helps, but it does not prove that a policy is still current. Recording the source version, modification date and status allows replaced or deleted passages to be removed. Permission changes must propagate to the index and caches within a defined window. If that window cannot be met for sensitive information, the retriever must consult current authorization or temporarily suspend that content.

Example: an assistant for return policies

A team starts with a bounded set of approved policies for two countries. Each document has a country, effective period and owner. For a query, the service identifies the user, filters their scope and retrieves the current policy. The answer includes the passage supporting the condition and the authorized link. When policies conflict, it asks for the country or escalates the query. This example is an implementation proposal, not a result attributed to Cloudata or a confirmed Dova feature.

During the pilot, the assistant explains procedures; it does not approve refunds or modify orders. Separating queries from execution makes it possible to evaluate knowledge quality before introducing actions with financial impact.

Measure more than an answer that sounds right

Build known, ambiguous, unanswerable and unauthorized questions. Keep test cases versioned and repeat them when documents, indexes, prompts or models change. Evaluations help check behavior against defined criteria [3]; acceptable limits should be agreed with the process owner.

Before opening access to more users

• Define eligible documents, an editorial owner and a correction mechanism.
• Test group changes, deleted documents and links that are no longer accessible.
• Review answers with conflicting sources and malicious content.
• Record evidence and evaluation results without unnecessarily storing secrets or complete documents.
• Agree on alerts, service withdrawal and a channel for reporting incorrect answers.

Dova is presented as a data integration platform with mappings, workflows, sessions, workers and monitoring. These capabilities belong to data flow preparation and operation. This RAG architecture requires additional components and its own validation; it does not assume that Dova includes a RAG engine or these authorization controls.