Overview
A fictional SaaS support workflow: index a policy corpus, select an approved knowledge set, retrieve bounded evidence, and draft a typed reply for review. The keyless default forces wrong-policy retrieval, checks context limits, skips generation when evidence is missing, and rejects invented source labels. Run from the repository root:--live uses OpenAI embeddings and generation on the same corpus. It requires
configured credentials. Fixture vectors and responses prove mechanics only;
the live run checks provider behavior separately. Neither mode sends a reply
or issues a refund. See the adjacent README for checks and observed output.