Help me choose a vector database for embeddings with metadata filtering and frequent updates

If your main requirements are:

• vector search over embeddings
• metadata filtering
• frequent updates/upserts
• likely production use

then the best choice depends mostly on your scale, latency needs, and how much operational complexity you want to manage.

Short recommendation

Best all-around choice for most teams: Postgres + pgvector

Choose this if:

• your dataset is not huge (roughly up to a few million vectors, sometimes more with tuning)
• you want strong metadata filtering
• you already use Postgres
• you need frequent updates and transactional consistency
• you want the simplest architecture

Why it fits:

• metadata filtering is excellent because it’s just SQL
• updates/inserts are straightforward
• operationally simpler than running a dedicated vector DB
• good for hybrid workloads where metadata queries matter a lot

Tradeoff:

• not the fastest at very large scale compared with specialized vector DBs
• ANN performance and tuning matter a lot

---

Best specialized vector DB for filtering + updates: Qdrant

Choose this if:

• you want a dedicated vector database
• metadata filtering is important
• you expect frequent upserts/deletes
• you want solid performance without too much complexity

Why it fits:

• strong payload/metadata filtering
• good support for upserts and deletes
• relatively easy to operate
• works well for dynamic datasets

Tradeoff:

• less flexible than SQL for complex relational queries
• you’ll manage a separate system from your primary DB

---

Best for high scale / heavy production search: Milvus

Choose this if:

• you expect large-scale vector workloads
• you need high throughput and scale
• you’re okay with more operational complexity

Why it fits:

• built for scale
• supports filtering and updates
• strong ecosystem for vector search

Tradeoff:

• more moving parts
• heavier to operate than Qdrant or pgvector

---

What I’d pick by scenario

1. Startup / MVP / product with rich metadata

Pick: Postgres + pgvector

• easiest to build
• no extra infrastructure if you already have Postgres
• excellent filtering and transactional updates

2. Need dedicated vector search, but want practicality

Pick: Qdrant

• probably the best balance of simplicity, filtering, and update friendliness

3. Very large scale or long-term search infrastructure

Pick: Milvus

• better when scale dominates everything else

---

Important considerations for your use case

Metadata filtering

If you need filters like:

• tenant_id = X
• status IN (...)
• created_at > ...
• tags contains ...

Then:

• Postgres is strongest and most flexible
• Qdrant is very good
• Milvus is good, but filtering semantics can be less convenient than SQL

Frequent updates

If vectors or metadata change often:

• prefer systems that handle upserts well
• avoid designs that require expensive reindexing or batch-only workflows

Good choices:

• Qdrant
• Postgres + pgvector
• Milvus if scale justifies it

Hybrid search

If you want both:

• keyword search
• metadata filters
• embeddings

Then:

• Postgres can be convenient for combining everything
• Elasticsearch/OpenSearch can work, but they’re usually less ideal than a dedicated vector DB for pure embeddings
• Qdrant integrates nicely with app-side ranking

---

Simple decision rule

• Use Postgres + pgvector if you want the simplest and most flexible solution.
• Use Qdrant if you want a dedicated vector DB with strong filtering and frequent updates.
• Use Milvus if you need maximum scale and can handle more ops complexity.

---

My default recommendation

For your stated needs, I’d usually recommend:

Qdrant if you want a dedicated vector DB, or Postgres + pgvector if you want the simplest architecture.

If you want, I can also give you:

1. a feature comparison table of Pinecone vs Qdrant vs Milvus vs pgvector, or
2. a recommendation based on your exact workload if you share:
   • vector count
   • embedding dimension
   • QPS
   • update rate
   • filter complexity
   • cloud/on-prem preference