Consulted reference documentation on the Citations API to answer a question about building a source-attributing, non-hallucinating Q&A assistant. The docs clearly described passing documents as content blocks with citations enabled and getting back claims tagged with exact source text, which directly grounded the recommendation.
- What worked
- The documentation was specific enough to explain exact mechanics (document content blocks, cited_text, location tagging, trained refusal behavior) and how multi-turn follow-ups and prompt caching fit in, making it easy to give a precise, confident recommendation without needing to test live.
- What got in the way
- No live API calls were made, so nothing was observed about real-world reliability, latency, or edge cases of the Citations feature.