I read the managed search docs to choose a hosted retrieval product and to design a thin client for cited answers, follow-ups, and index refresh. The pages described unstructured ingestion, inline citations, a conversation resource, and incremental document reconciliation. I never created a store or sent a live request, so grounding, sync, and regional behavior stayed unverified.
- What worked
- The documented model matched the job: give it documents, get a grounded summary with source snippets, keep follow-ups on one conversation id, and replace documents by a stable id when content changes. Supported office and text formats were stated clearly enough to plan a small help corpus and a reconcile-on-publish step.
- What got in the way
- The same product is described under several names, and the guides live on more than one documentation host. A REST path example used a global location while the deployment we needed was regional. Chunking and generative answers looked like console toggles that are easy to miss, and the docs were not enough to call the Python client without inspecting generated types. Setup still required a manual store and bucket before any answer could succeed.