Twelve questions that predict whether a retrieval-augmented AI pilot survives production. None of them are about the model. When these systems fail, the retrieval is almost always the problem, and retrieval quality is a product problem, not an infrastructure problem.
Answer honestly. "We planned to" counts as a red flag.
What the system is allowed to know. Most failures are decided here, before a single query is run.
Who owns the documents the system retrieves from, and when were they last updated?
Stale and contradictory sources produce confidently wrong answers. Someone has to own freshness.
Red flag: "We indexed everything in the shared drive."
Has a human actually read a sample of the chunks?
A chunk that cuts a policy in half retrieves half a policy. Default splitting settings are a guess, not a decision.
Red flag: Default chunk size, never inspected.
What must the system never retrieve, and who enforces that?
Access control lives in the retrieval layer. If confidential documents are in the index, they are in the answers.
Red flag: Permissions were "planned for a later phase."
Whether the right passages reach the model. If this step fails, nothing downstream can save the answer.
Do you evaluate retrieval separately from generation?
If you only judge the final answer, you cannot tell whether the model failed or never saw the right passage.
Red flag: Quality is checked by reading answers and nodding.
What happens on exact-term queries: product codes, names, abbreviations?
Pure semantic search misses exact terms. Hybrid search exists because real users type part numbers, not paragraphs.
Red flag: Nobody has tested a query with a code in it.
How many test questions exist, and who wrote them?
A test set written by the team that built the system tests the team's imagination, not the users' behavior.
Red flag: Five demo questions, all written in-house.
Whether the answer stays honest to its sources, and what happens when there is no source.
Can every answer point to the passage it came from?
Citations are the audit trail. Without them, wrong answers are indistinguishable from right ones.
Red flag: Citations exist but nobody clicks them.
What does the system do when the retrieved passages do not contain the answer?
The most expensive failure mode is a fluent answer grounded in nothing. "I don't know" must be a designed behavior.
Red flag: It answers anyway.
Has anyone measured whether answers stay faithful to the retrieved text?
A model can retrieve the right passage and still drift away from it. Faithfulness is measurable. Measure it.
Red flag: "The answers look right to us."
Week three and beyond. This is where demo luck runs out.
Does anyone read the query logs?
Real users never ask the demo questions. The logs are the only honest test set you will ever get.
Red flag: Logs exist, unread.
Who owns answer quality as a metric?
Engineering owns the pipeline. If nobody owns the answers, quality decays quietly and everyone is surprised later.
Red flag: Silence, then "the team, collectively."
What is the plan when the documents change?
Policies get revised, products get renamed, prices move. An index without a reindexing cadence is a countdown.
Red flag: "We'll reindex when we notice a problem."
None of them require you to build anything, and none of them are about the model. They are ownership questions. A RAG system rarely fails loudly; it fails in the gap between the team that built the pipeline and the person accountable for the answers.
If your red flag count worries you, that is exactly what a Diagnostic Day is for: one day, your system, these questions asked properly, and a written report of where the risk sits.
Book a Diagnostic Day →