LLM evaluation in production

The demo convinces. Two weeks later someone changes the model. Answers get shakier. Without evaluation nobody notices in time.

What evaluation means here

Not a one-off blind test with five favourite questions. A fixed set: question, expected source or decision, who maintains the set. After every prompt, model or index change the same set runs.

Without that, "better" is a feeling. In an audit a feeling counts for nothing.

What I measure

Does the answer hit a source a human can open?

Does the system say no when the base is missing?

Does the hallucination rate hold after a model change?

Does cost rise without quality rising?

Tools like RAGAS help. They do not replace the rule. The rule sits in the flow: measure, then change. Not the other way around.

A typical finding

A team switches to a larger model. The prose sounds better. The source rate drops. Legal would have seen that only at acceptance. With a set of 40 questions it would have been visible on the day of the switch.

This hangs on source grounding. Evaluation without a source checks style. Evaluation with a source checks whether you may go live.

What the readiness check does with it

On day 3 of the LLM readiness check I look for a set and for someone who runs it after a change. If it is missing, the report calls it a blocker, not a nice-to-have.

Related: reconstructing answers, hybrid RAG. The service page is AI engineering. Proof from practice: the audit assistant.