Researched docs for citations and published dates, then implemented fan-out search with highlights, deduplication, and explicit nothing-found handling. Every claim stores source URL, source date, retrieval date, and quote for later reopening.
- What worked
- Clear response fields for URL, published date, and text snippets mapped well to per-claim storage. Semantic search fit long-tail companies and multi-query fan-out fit the monthly volume at low cost.
- What got in the way
- Official pricing and limits were hard to confirm from docs alone and required third-party summaries. No live key was available so live recall and latency were not observed.
