I'm building a RAG pipeline and want a reliable vector store

A reliable vector store for RAG usually means one that is:

• Stable and durable: persists data safely, supports backups/replication
• Fast at retrieval: low-latency similarity search at your scale
• Operationally manageable: easy to deploy, monitor, and maintain
• Feature-complete for RAG: metadata filtering, hybrid search, upserts, deletes, namespaces/collections

Good options by scenario

1) Best all-around managed option

Pinecone

• Strong reliability and simple ops
• Good performance and scaling
• Easy metadata filtering
• Good if you want minimal infrastructure work

2) Best open-source/self-hosted options

Qdrant

• Very solid reliability
• Great filtering and payload support
• Easier to operate than some alternatives
• Good choice if you want self-hosted but production-ready

Weaviate

• Strong feature set, hybrid search, metadata filtering
• Good for more complex retrieval setups
• Heavier than Qdrant operationally, but capable

3) If you already use a traditional DB

Postgres + pgvector

• Reliable and familiar
• Great for smaller to medium workloads
• Simplifies architecture if you already run Postgres
• Can become limiting for very large-scale, high-throughput vector search

4) If you need high-performance local/self-managed search

Milvus

• Powerful at scale
• More operational complexity
• Good when you need large-scale vector workloads and can handle infra

My practical recommendation

• If you want the most reliable and easiest path: Pinecone
• If you want self-hosted and production-friendly: Qdrant
• If your scale is modest and you want simplicity: Postgres + pgvector

What to evaluate before choosing

1. Persistence and backup strategy
2. Replication / HA support
3. Metadata filtering
4. Upsert/delete behavior
5. Latency under your expected QPS
6. Index build time and rebuild strategy
7. Support for hybrid search if you want lexical + vector retrieval
8. Cost at your expected scale

Simple rule of thumb

• Prototype / small app: pgvector
• Production with low ops burden: Pinecone
• Production self-hosted: Qdrant
• Large-scale / advanced infra: Milvus

If you want, I can help you choose one based on your:

• expected document count
• query volume
• cloud/on-prem preference
• need for metadata filtering or hybrid search