An indexer can finish its run while individual documents fail. If the allowed failure count is nonzero, “success” may mean the pipeline continued, not that every source object became searchable. Azure AI Search's indexer monitoring documentation explicitly distinguishes overall run success from per-document errors. The failure might be a parsing problem, a new field type the index cannot accept, an oversized document or a permission problem.

Start with one missing source ID. Follow it through discovery, extraction, transformation and index write. Did the source version enter the run? Was it rejected with an error tied to that document key? Is an older version still searchable? A stale old page is more dangerous than an obvious gap if the assistant confidently cites it as current policy.

I would define the success invariant at the document level. For each source version intended for publication, record a terminal outcome: indexed at version V, deliberately excluded with a reason, or failed and awaiting repair. Compare source inventory with indexed version inventory by tenant and document family. Alert on failure counts and age, not just job status or consumer lag. Keep rejected payload references and error classes in a repair queue, with access controls, so an operator can replay after fixing the parser or mapping. A retry must be idempotent and must not let an old failure overwrite a newer indexed version.

The failure policy depends on the use case. A public knowledge article may tolerate a small, visible delay. A new refund-policy exception that changes customer eligibility may require a publication barrier or temporarily suppressing answers in that scope until the index catches up. That is a product and source-authority decision, not a generic “always fail the entire batch” rule.

When the interviewer says the queue lag is zero, I would ask which acknowledgments that metric counts. A source event can be consumed and the indexing run can finish while one document is rejected. The useful evidence is end-to-end source-to-search coverage, including the exact versions users can retrieve.