Grounded-Answer Systems Fail at Retrieval, Not at Generation

Systems that answer questions from an organisation's own documents work by retrieving relevant passages and asking a language model to answer using them. When such a system produces a wrong answer, the failure is usually reported as a model failure — it made something up, it misunderstood, it needs a better model. Teams that instrument the pipeline properly find a different distribution: in most wrong answers, the passages handed to the model did not contain the correct information, and the model's output was a faithful summary of inadequate input. The retrieval failures cluster into recognisable shapes. The corpus does not contain the answer, and nothing in the pipeline is able to say so, producing a confident answer assembled from adjacent material. The answer exists but is phrased with none of the query's vocabulary, so similarity search never surfaces it. The document is present in three versions and the retrieved one is obsolete, which is the most damaging case because the answer is well-formed, well-sourced and wrong. Or the answer requires combining two documents that are individually only weakly relevant, and a ranked list of individually-relevant passages never assembles them. None of those are improved by a more capable model, and some are made worse: a stronger model produces a more fluent and more convincing answer from the same bad passages. The improvements that move the metric are unglamorous — deduplicating and versioning the corpus so there is one current answer, enriching documents with the vocabulary users actually type, combining keyword and semantic matching rather than choosing one, and, most importantly, building an honest empty state so the system can say the corpus does not cover this. That last capability is the one most often missing, and it is the difference between a tool people trust and a tool people verify. A system that answers everything provides no signal about which answers to check, which means every answer must be checked, which means the system has saved nobody any time. A system that declines the questions it cannot ground is less impressive in a demonstration and considerably more useful on a Tuesday.