Integrated a managed serverless vector index with hosted inference embeddings to match meaning without self-hosting, keyed by part identifier with lazy loading and degraded UI when unconfigured. Verified locally with unit tests using fakes and client signature checks; live index creation and backfill were not run for lack of key and network.
- What worked
- One key for embeddings plus storage avoided a second vendor, lazy imports kept the app runnable without the package or key, and explicit unavailable messaging avoided silent empty results.
- What got in the way
- The live service was never contacted here, so end-to-end similarity quality, latency, and cost remain unverified and response shapes needed defensive accessors.
