I'm building an LLM app and need a vector store with metadata filters

If you need a vector store for an LLM app with metadata filtering, here are solid options depending on your scale, ops preferences, and budget:

Good choices

1) Postgres + pgvector

Best for: MVPs, moderate scale, apps that already use Postgres
Why:

• Supports vector search plus SQL metadata filters
• Simple operationally if you already run Postgres
• Easy to combine structured queries with embeddings

Example use case:

• WHERE tenant_id = ... AND doc_type = ... AND updated_at > ...

Tradeoff:

• Not as fast/scalable as dedicated vector DBs at very large scale

---

2) Pinecone

Best for: Production apps that want managed scaling
Why:

• Built-in metadata filtering
• Strong performance and operational simplicity
• Good for high-concurrency retrieval

Tradeoff:

• Managed service cost
• Less flexible than self-hosted options

---

3) Weaviate

Best for: Teams wanting rich filtering + hybrid search
Why:

• Good metadata filters
• Supports hybrid lexical + vector retrieval
• Schema-driven, feature-rich

Tradeoff:

• More moving parts than pgvector
• Operational overhead if self-hosted

---

4) Qdrant

Best for: Fast filtering-heavy workloads
Why:

• Excellent payload/metadata filtering
• Strong performance
• Flexible and developer-friendly

Tradeoff:

• Another service to manage if self-hosted, though cloud is available

---

5) Milvus

Best for: Very large-scale vector workloads
Why:

• Built for scale
• Supports filtering
• Good for heavy retrieval systems

Tradeoff:

• More complex operationally

---

Quick recommendation

• Starting out / already on Postgres: pgvector
• Want managed + easy scaling: Pinecone
• Need powerful filtering and good performance: Qdrant
• Want hybrid search features: Weaviate

What to look for

Make sure the store supports:

• Metadata filters on fields like tenant_id, doc_type, source, timestamp
• Hybrid search if you need keyword + semantic
• Upserts and partial updates
• Multi-tenancy isolation
• ANN indexing appropriate for your scale

Example filter patterns you may want

• Tenant isolation: tenant_id = "acme"
• Document type: doc_type IN ["policy", "faq"]
• Time constraints: created_at >= "2026-01-01"
• Access control: allowed_group_ids contains "sales"

If you want, I can also give you:

1. a decision matrix,
2. a recommended architecture, or
3. code examples for LangChain / LlamaIndex / raw SDKs.