Because the arithmetic usually removes the question. 650,000 chunks come to about 4 GB, which fits in memory on one ordinary machine — so distributed systems, sharding, and the databases built for them are all irrelevant. Sizing first turns an open-ended technology debate into a short list. It also exposes the cases where you genuinely do need heavy infrastructure, which is far rarer than people expect.

Because it deliberately answers one question with the answer to a similar one. In support, a near-miss produces a slightly-off article and a mildly annoyed customer. In clinical use, "30 mg" and "300 mg" embed almost identically while having opposite safe answers, so the same mechanism can produce a dosing error. The technique did not change; the cost of its failure mode did.

Because it is counted automatically for every interaction, and it maps directly onto money. Satisfaction scores depend on users choosing to respond, which a small and unrepresentative group does, and they drift with mood and wording. Escalations are recorded whether or not anyone opts in, they are unambiguous, and improving them has a value the business already knows how to price.

Because Stage 1 produces the baseline that makes every later decision checkable. Without it you cannot show that hybrid search helped, cannot tell how much reranking is worth, and cannot spot the change that quietly made things worse. It also forces the golden dataset into existence early. Building the full pipeline first leaves you unable to remove anything, since nothing was ever measured on its own.

Careful parsing, sensible chunking, hybrid search, and reranking. These serve almost every system because they address problems every corpus has — mangled documents, boundaries splitting facts, exact terms that embeddings miss, and rough initial ranking. What varies is the layer above: caching policy, refusal behaviour, verification depth, permissions, and oversight, all driven by what a wrong answer costs.