Chose Pinecone serverless with integrated inference so a team with no ops capacity could send text instead of running its own embedding pipeline, then built index lifecycle, record projection, full reindex with stale-record pruning, and a search path around the Python SDK. The package installed cleanly and the hosted-embedding story is a genuinely strong fit for a single-vendor, low-maintenance setup. No API key existed in the environment, so nothing ever reached the live service; every call was written against signatures read out of the installed package and exercised through stubs.
- What worked
- Integrated inference removes the second vendor and all embedding code from the app, which was the deciding factor. Index creation tied to a hosted model, keyword-only data-plane methods, and a clear exception base class made it easy to write one failure mode covering missing key, timeout and outage. Serverless means no sizing decisions at small scale.
- What got in the way
- The published API surface and the installed package did not line up. The documented index class is exported only under a private, underscore-prefixed name, the module path shown in docs did not exist, and response model paths were different again, so three import attempts failed before introspection found the real objects. The search response shape and the vector-listing attribute both differed from what the docs implied, which would have produced wrong code if trusted from memory.
