Index Management

Operator must choose between serverless and pod-based for a workload with spiky read traffic, ~10M vectors at 1536 dim, and unpredictable namespace count.

Pick serverless via spec.serverless{cloud, region}: auto-scales, pays per read/write/storage unit, supports unbounded namespaces. Pod-based (p1/p2/s1) fits steady QPS with predictable size — do not default to pods just because they look 'production'. Document the dimensional and metric choice (immu…

Operator creates an index with metric=euclidean for vectors produced by an OpenAI text-embedding-3-small model (normalized — cosine-appropriate). Recall is poor.

Pinecone index metric is set at create time and immutable. The fix is to create a NEW index with metric=cosine (or dotproduct on normalized vectors) and reupsert, not to mutate. Verify embedder docs for the expected metric before create.

Operator creates a pod-based index then immediately issues an upsert. The upsert 404s.

After create_index, poll describe_index until status.ready=true (and status.state=Ready) before issuing data-plane calls. Pod-based indexes can take minutes to provision; serverless is faster but still not immediate. Use exponential backoff.