I'm unhappy with pgvector for large-scale embeddings

Totally fair — pgvector is convenient, but it’s not usually the best choice once you get into large-scale embedding workloads.

Common pain points:

• Index build / rebuild times get painful as vectors grow
• Query latency can become inconsistent at higher scale
• Postgres operational overhead rises because you’re forcing a search workload into an OLTP database
• Memory / vacuum / bloat / replication concerns can show up
• Tuning HNSW / IVFFlat in Postgres is less flexible than dedicated vector systems

If your goal is large-scale vector search, better options are usually:

Managed vector databases

• Pinecone – simplest for production semantic search
• Weaviate – good hybrid search + filtering
• Qdrant – strong OSS + managed option, good filtering
• Milvus – built for scale, common for very large deployments

Search engines with vector support

• OpenSearch / Elasticsearch – useful if you already need keyword + vector hybrid search
• Good when retrieval needs combine structured filters, full-text, and vectors

Specialized infra

• FAISS – great for DIY / offline / custom systems
• Often used when you can manage your own indexing and serving layer

When pgvector is still okay

It’s fine if:

• your corpus is relatively small or medium-sized
• you want simplicity and everything in Postgres
• you need transactional consistency tightly coupled with app data
• you’re still prototyping

A practical rule of thumb

If:

• you’re over hundreds of thousands to millions of vectors,
• need low-latency search,
• and vector search is becoming a core workload,

then moving off pgvector is often the right call.

A good migration pattern

A lot of teams do:

1. Keep Postgres as the source of truth
2. Replicate embeddings to a dedicated vector store
3. Use Postgres for metadata and joins
4. Use the vector DB for ANN retrieval
5. Optionally rerank in app or search layer

If you want, I can help you choose between Pinecone vs Qdrant vs Milvus vs OpenSearch based on:

• corpus size
• QPS
• filtering needs
• budget
• cloud/on-prem requirements