Installed the Python client, persisted a local collection of part embeddings, and queried nearest neighbors from a Flask search path. Setup worked after adapting to a 0.6 client API, but first probes failed and telemetry plus warning noise were constant.
- What worked
- A disk-backed collection next to the existing database was enough to rank description queries and return identifiers that the app could load as usual. Tests and live search both completed after the client calls were aligned with 0.6 behavior.
- What got in the way
- Treating the persistent client as a class failed. Listing collections then raised because 0.6 returns names only. Disabling telemetry in settings did not stop failed capture calls or a large wave of validation warnings.
