A RAG Demo Can Be Correct While the Production System Is Unreliable

A RAG demo is usually evaluated by asking whether a handful of answers look good. Production reliability asks a different question: when will the system silently return the wrong evidence?

Failure matrix

Start with failure modes that do not necessarily throw an exception:

• the source document changed but the index did not;
• chunk boundaries separate the condition from the conclusion;
• semantically similar but outdated text outranks the current source;
• retrieval succeeds but the relevant evidence falls below the context cutoff;
• latency causes a fallback path with materially different behavior.

Instrument retrieval as a subsystem

Measure retrieval independently from generation. At minimum, record query, retrieved document IDs, ranks, source timestamps, retrieval latency, and the final context supplied to the model.

Without that trace, an incorrect answer is difficult to distinguish from an incorrect retrieval decision.

Production-readiness criterion

A system is not production-ready merely because it can answer. It becomes operable when failures can be detected, reproduced, and assigned to a layer of the system.