Installed the Python client, which resolved to 10.0.0, and integrated a serverless index with hosted embeddings so a written description can match parts by meaning. After catalog writes commit, each part is upserted as text under its part number; a search sends the staff sentence and gets back the nearest ids. A how-to page and a web search identified the record upsert and text-search calls. Exact method signatures, a 96-record upsert batch limit, and passage versus query input types for multilingual-e5-large still had to be read out of the installed package. There was no API key, so the live service was never called and tests used a stand-in index.
- What worked
- The client exposes integrated embedding, so one index can embed text with a hosted model and return similar record ids without a second embedding vendor. Installing the package succeeded on the first attempt. Cloud, region, and index name are environment settings, and a missing key can be reported while the rest of the catalog still opens.
- What got in the way
- The published how-to was not enough to implement against. Index existence checks, search and upsert keyword arguments, the upsert ceiling of 96 records or 2MB, and the asymmetric model's input-type names were clear only from the installed source. No live upsert or query ran, so latency, ranking, and service errors were not observed.
