Chose this Search HTTP API as the gathering layer and implemented a client that maps ranked hits into append-only rows with passage, source URL, and retrieval time. Account, billing, and terms were taken from vendor docs only; tests used a fake HTTP adapter and the live service was never called.
- What worked
- Documented result shape (URL, title, publish date, excerpts, search id) matched the evidence needed on the claim file. Quickstart and pricing pages were enough to wire the request, pick advanced mode, and estimate cost at about forty named-firm claims a week.
- What got in the way
- Customer terms and the FAQ disagreed on training and reuse of customer data, so production still needs a DPA and queries limited to firm names. Pricing and credit rules were spread across several pages. Hosted reliability was not observed because no key was issued.