Why Semantic Search Can't Answer 'What's Our Total Contract Exposure?'
Document-chat tools mostly rest on Retrieval Augmented Generation, and part of that machinery is semantic search: at query time the system pulls back only the top-k chunks of text most relevant to a question. AWS's machine learning blog points at the question shape where that design gives out.
Ask something that spans every contract in a set — total exposure, or which agreements are about to expire — and the answer has to be aggregated across the full dataset, not assembled from a handful of retrieved chunks. Retrieval answers questions whose answer lives somewhere; aggregation questions have no single somewhere.
The blog's conclusion is unusually blunt for vendor documentation: a better prompt won't fix this, a different architecture will. That is a useful line to carry into any demo — some failures are prompt failures, and some are architecture failures wearing the same costume.
AI-generated discussion · Different perspectives
Worth noticing who is making the diagnosis. A vendor arguing that retrieval can't do aggregation is also arguing for the different architecture it happens to sell.
The incentive doesn't dissolve the distinction. Pulling the most relevant passages and summing across an entire dataset are genuinely different operations — one is a search problem, the other is a counting problem.