A RAG demo can look convincing with ten documents and prepared questions. Production begins when the collection changes, permissions matter, questions are ambiguous and a wrong answer costs time or trust.
The system should not be evaluated only by how fluently it writes. It must prove that it finds relevant evidence, respects access, recognizes missing information and lets the team reconstruct how an answer was produced.
Separate retrieval from generation
When an answer fails, the team needs to know whether the source, indexing, query, ranking or generative model caused it. Measuring only the final answer hides that chain. A useful evaluation keeps expected fragments and checks whether they appear with sufficient priority.
Versions matter too: document, indexing date, retriever configuration and model. Without them, comparing an improvement with previous behavior becomes an impression.
Permissions belong inside retrieval
Filtering after generation is too late. Identity, tenant, role and document classification must constrain which fragments can enter context. Tests should include attempts to cross clients, departments and confidentiality levels.
The principle is simple: an answer must not reveal or infer content the user could not open in the original source.
Visible evidence and abstention
A useful citation points to a specific fragment and document, not a generic source list. Users should be able to open the context and verify whether it supports the claim. When evidence is insufficient or contradictory, not answering and proposing search or human escalation may be the correct behavior.
- Show relevant fragments and metadata.
- Separate answer, interpretation and lack of evidence.
- Flag obsolete or conflicting documents.
- Prevent instructions inside documents from governing the system.
Evaluate with real questions
The evaluation set should cover easy, ambiguous, unanswered, sensitive and adversarial questions. Retrieval, faithfulness, usefulness and safety should be reviewed separately. Automated metrics detect regressions, but a human sample remains necessary to decide whether the system supports the process.
Before expanding users, an agreed threshold and incident procedure should exist: which logs are retained, who reviews an answer, how a source is corrected and when a capability is disabled.
The exit criterion
RAG is ready when the team can explain which corpus it uses, who may see it, how it is updated, which questions it resolves, how it is measured and what happens when evidence is insufficient. Latency and cost matter, but they do not replace those guarantees.
Frequently asked questions
Does a citation automatically make an answer reliable?
No. The citation must be relevant and support the statement. A related but insufficient document only creates the appearance of rigor.
Can RAG be evaluated only with an LLM judge?
It can accelerate analysis, but it should be calibrated with human review and deterministic tests for permissions, retrieval and format.
What should happen when evidence is missing?
The system should abstain, say so clearly and offer a safe route: reformulate, search another source or escalate to a person.
