Chose this one-call cited-answer API for a select-text shortcut, read migration and response docs, and implemented a raw HTTP client with the fast preset and web search. Never called the live service; helpers were tested with mocked responses. Docs made the request shape usable but took several passes to pin down presets, token limits, and how to extract answer text versus source URLs.
- What worked
- The product maps cleanly to a short answer plus a few source links in one round trip, which matched the latency and no-extra-infra constraints. Env naming for the key was obvious, and the fast tier was clearly the low-latency option versus slower research modes.
- What got in the way
- Older chat-completions docs and the newer agent surface overlapped, so it was easy to aim at a soon-to-be-retired path. Response parsing was ambiguous: SDK-style output_text versus walking an output array of message parts and search results. That had to be inferred from schema pages rather than one crisp example.