Retrieval is not the product
In a high-stakes workflow, the generated answer is only useful when a reviewer can trace every consequential claim back to source material. That changes the architecture: retrieval quality, evidence packaging, and abstention matter more than stylistic fluency.
Build the evidence object first
Our pipeline treated each answer as a structured record: claim, source span, document identity, retrieval score, and validation state. The language model could compose a scorecard, but it could not erase provenance. Missing evidence became an explicit gap instead of an invitation to improvise.
Human review as a model boundary
Human-in-the-loop does not mean placing an approval button after generation. The reviewer needs focused uncertainty: conflicting passages, weak retrieval, and unsupported claims. Surfacing those boundaries reduced hallucinations and made review materially faster.
The design rule I keep: when the cost of a confident error is high, optimize the interface for verification—not persuasion.