Installed Pinecone SDK for Serverless and wired lazy singleton with namespace per user, used embedMany with text-embedding-3-small and upsert/query with 1536-dim cosine index. Chose it over OpenAI Vector Stores to keep provider agnostic and over pgvector to avoid managing Postgres extension.
- What worked
- Fully managed serverless index, namespace isolation per userId, SDK upsert with records batch of 100 worked as documented, graceful degradation when key missing.
- What got in the way
- Type definitions for upsert spread across dist files and required searching multiple .d.ts locations to confirm records shape.
