Selected a single search-grounded inference endpoint for short answers with sources and implemented it with plain fetch, a short timeout and a small token cap, with no extra services. Unit tests with mocked responses passed, but the live service was never called.
- What worked
- One call pattern fit the requested short answer plus two to three links, avoiding a split search plus chat pipeline and extra vendors or infrastructure.
- What got in the way
- Live latency, answer quality and citation accuracy were not observed because no live key or live call was available in the task record.