AMA answered a spend question with $193.95 from 8 charges when 40 receipts — 37 in USD (US$583.06), 3 in CA$ (CA$357.74) — were stored, visible, and already embedded — it summed a 30-row sample and returned the sample's total as if it were the whole answer. There is no single correct total across those 40 without an FX conversion, which is exactly why a `$` sitting in unstructured text is not a currency: the amount has to be a typed field.
The governing split. A RECALL question ("what did Regina say about the lease") needs relevance over a large corpus and is correctly served by top-k retrieval over embedded documents — a sample is the right tool, because the answer is "the most relevant few," not "all of them." An AGGREGATION question ("how much did I spend," "how many receipts," "list every invoice") needs completeness over the full matching set, not relevance — the correct answer is a property of every row that matches, and top-k over embeddings will silently under-count no matter how the prompt is worded. Serving an aggregation question through the recall lane is the defect; the fix is routing aggregation questions to an exact query lane instead of asking retrieval to do a job it cannot do.
Tracked in the epic. See the issue for the current write-up; this page does not restate a body the issue already carries and would drift from.
Tracked in the epic. See the issue for the current write-up.
Tracked in the epic. See the issue for the current write-up.
The three p0s come first because they are the defect a user already hit — a wrong number stated with confidence, not a missing feature. The three p1s follow because each is a precondition for the fix that closes the p0s: an aggregation question cannot be answered completely by a retrieval lane built for relevance, so the exact-query lane has to exist before the routing decision has anywhere correct to send the question.
What this order deliberately does NOT use:
#4267, #4268, #4271. A closed issue named as current work is the drift this section exists to surface — re-derive the prose, or move the reference into History.
Evidence below is generated from the read-only audit of 2026-08-29 and reconciled against the open issues at build time. Every claim cites the issue it came from. Nothing here is a decision — rulings stay in the prose above, which is written by hand.
Filed as the epic #4266 after AMA answered a spend question with $193.95 from an 8-charge sample while 40 receipts — 37 USD totalling US$583.06, 3 CA$ totalling CA$357.74 — were stored, visible, and already embedded — the answer summed a 30-row sample and returned it as if it were complete. Three p0 children capture the defect; three p1 children capture the exact-query lane the fix needs.