I'm unhappy with pgvector because I need better filtering and scale

If you’re unhappy with pgvector because of filtering + scale, that usually means you’ve hit the point where “vectors inside Postgres” is no longer the best fit.

Common reasons pgvector starts hurting

• Complex metadata filtering becomes slow or awkward at high cardinality
• Index performance degrades as the collection grows
• Hybrid search (vector + keyword + facets + recency) gets cumbersome
• Operational load competes with your primary OLTP database
• Relevance tuning becomes limited compared with dedicated vector/search systems

Better alternatives, depending on what you need

1) Dedicated vector database

Best if your main pain is ANN search at scale.

• Pinecone – managed, easy scaling, strong filtering
• Weaviate – flexible schema, hybrid search, decent filtering
• Qdrant – strong filtering, good performance, open source + managed options
• Milvus – good for very large scale, more operational overhead

Good choice if: vector retrieval is a first-class workload and filtering must be fast.

2) Search engine with vectors

Best if filtering + keyword + vector hybrid search matters.

• Elasticsearch / OpenSearch
• Good at:
  • structured filters
  • text search
  • ranking
  • faceting/aggregations
  • vector search “good enough” for many apps

Good choice if: you need filters and search UX more than pure vector DB features.

3) Separate systems

A common architecture:

• Postgres for transactional metadata
• OpenSearch/Elastic or a vector DB for retrieval
• sync via CDC / queue / background jobs

This often gives the best balance of correctness and scale.

What to look for in the replacement

If filtering is the blocker, prioritize:

• Pre-filtering support at index/query time
• High-cardinality filter performance
• Hybrid retrieval
• Sharding / horizontal scaling
• Metadata + vector in the same query
• Latency under filtered ANN queries

Practical recommendation

• If you want best filtering + search UX: try OpenSearch/Elasticsearch
• If you want best vector-first filtering: try Qdrant or Pinecone
• If you expect very large scale and can manage ops: Milvus

Migration pattern

1. Keep Postgres as source of truth
2. Export embeddings + metadata to new engine
3. Dual-write for a while
4. Compare recall/latency/filter correctness
5. Cut over reads once stable

If you want, I can help you choose among Qdrant vs Pinecone vs OpenSearch vs Weaviate based on:

• your data size
• filter complexity
• update rate
• latency target
• budget
• cloud/on-prem preference