Our documentation changes every sprint and the assistan...

To solve this issue, analyze the update shape: a small fraction of a fixed corpus changing on a recurring two-week cadence. Choosing fine-tuning to fix deprecated documentation is ineffective because fine-tuning pays the computational cost of a full training run to move a tiny fraction of the corpus. Furthermore, fine-tuning requires replay data to prevent catastrophic forgetting on the remaining content and still does not reliably guarantee verbatim endpoint citations.

A retrieval-augmented approach scales directly with the update shape. When endpoints change, only the affected documentation chunks in the retrieval index need to be updated or re-embedded. This minimizes compute costs, guarantees that the model receives the exact updated context during inference, and directly solves the problem of citing deprecated endpoints.