Installed the Python client, inspected conversational search, document, schema, and data-store APIs locally, then implemented chat, upsert, delete, and store setup against those types with tests fully mocked. Install and import were smooth; GAPIC flattened-field rules were easy to get wrong until signatures were checked.
- What worked
- After install, imports and path helpers were consistent. Inspecting client signatures made it possible to correct request objects for schema and document updates and to see which converse fields are flattened. Summary, reference, and citation types were present and usable for parsing sources.
- What got in the way
- Several mutating calls require a request message rather than the keyword arguments the first implementation used; converse also does not flatten serving config or filter. An inspection script hit a bound-method quirk. No RPCs were sent, so live error quality is unknown.