OpenAI

Vectors from different embedding models are not comparable, so the whole corpus must be re-embedded. Build the new index alongside the old one, keep queries on the old index until the new one is complete and evaluated, then cut over and only afterwards delete the old vectors.

Build a unit-cost model from a pilot: measured p50 and p95 input and output tokens per request, multiplied by the model's per-million rates and forecast volume. Then enforce it with per-project budgets, separate keys per feature for attribution, and alerts on tokens per request.

Skip strict schemas when the output is prose, when the shape is decided per request so every call pays fresh grammar compilation, when the schema cannot fit the restricted dialect or its size caps, or when portability across providers outweighs the guarantee.

The tool runs model-written Python in a sandboxed container with no open internet access. The container is short-lived, expiring after a period of inactivity, is billed per session on top of tokens, and any files it produces are lost unless you download them first.

Tool definitions are serialized into the prompt and billed as input tokens on every request, and selection accuracy falls as similar tools multiply. Keep the per-request set small and distinct; route or split across sub-agents rather than growing one surface.