I'm building a self-hosted vector search stack, what are the best options?

If you want a self-hosted vector search stack, the “best” option depends mostly on whether you want:

1. a purpose-built vector database,
2. vector search inside an existing database, or
3. a search engine with vector support.

Here are the strongest choices by category.

---

Best self-hosted vector DB options

1) Qdrant

Best for: most production teams wanting a clean, easy-to-run vector database

Why it’s great

• Simple architecture and API
• Strong filtering + payload support
• Good performance for ANN search
• Easy to self-host with Docker/Kubernetes
• Good balance of features and operational simplicity

Tradeoffs

• Smaller ecosystem than Postgres/OpenSearch
• Less “general search engine” functionality than OpenSearch/Elastic

Choose Qdrant if: you want a dedicated vector DB that is straightforward and production-friendly.

---

2) Milvus

Best for: large-scale vector search and heavy throughput

Why it’s great

• Very popular at scale
• Designed specifically for vector workloads
• Rich indexing options and distributed architecture
• Good fit for very large datasets

Tradeoffs

• More complex to operate
• Heavier stack than Qdrant
• Often overkill unless you need scale

Choose Milvus if: you expect big data, distributed deployment, or high QPS and are willing to manage more complexity.

---

3) PostgreSQL + pgvector

Best for: teams already using Postgres, smaller-to-medium scale, or mixed relational + vector workloads

Why it’s great

• One database for metadata + vectors
• Easy operationally if you already run Postgres
• Great for hybrid app logic, filters, transactions
• Strong developer familiarity

Tradeoffs

• Not as fast or specialized as dedicated vector DBs at large scale
• Can become limiting for very large collections or very high query rates

Choose pgvector if: you want the simplest path and your scale is moderate.

---

Best search-engine style options

4) OpenSearch / Elasticsearch

Best for: hybrid search, keyword + vector search, enterprise search use cases

Why it’s great

• Excellent full-text search
• Hybrid retrieval with BM25 + vectors
• Filtering, scoring, aggregations, observability
• Mature ecosystem

Tradeoffs

• More operational overhead
• Vector search is not always as specialized as dedicated vector DBs
• Can be expensive in resources

Choose OpenSearch/Elasticsearch if: you need classic search capabilities along with vectors.

---

Other notable options

5) Weaviate

Best for: developer-friendly semantic search with rich features

Why it’s great

• Good vector search experience
• Hybrid search support
• Easy schema and API patterns
• Useful modules/ecosystem

Tradeoffs

• Operational complexity varies by setup
• Some teams prefer the simplicity of Qdrant

Choose Weaviate if: you like its feature set and API style.

---

6) Vespa

Best for: advanced ranking, large-scale search, custom retrieval pipelines

Why it’s great

• Powerful ranking and retrieval capabilities
• Combines vectors, text, and structured data well
• Highly customizable

Tradeoffs

• Steeper learning curve
• More complex to deploy and tune

Choose Vespa if: you need sophisticated ranking/search and can handle complexity.

---

Quick recommendations by use case

If you want the simplest good production choice:

• Qdrant

If you already use Postgres and want to keep it simple:

• Postgres + pgvector

If you need best hybrid text + vector search:

• OpenSearch or Elasticsearch

If you need large-scale distributed vector search:

• Milvus

If you want advanced retrieval/ranking pipelines:

• Vespa

---

A practical stack recommendation

For many self-hosted applications, a strong modern setup is:

• Embeddings model: local or hosted model
• Vector store: Qdrant
• Metadata store: PostgreSQL if needed
• App cache/queue: Redis
• Ingestion pipeline: your ETL / workers
• Optional reranker: cross-encoder or LLM reranker

If you need keyword search too:

• Qdrant + OpenSearch or just OpenSearch for a unified search layer

---

How to choose in 30 seconds

Ask:

• Do I need full-text search too?
  → OpenSearch/Elasticsearch

• Do I want simplest dedicated vector DB?
  → Qdrant

• Do I already run Postgres and my scale is moderate?
  → pgvector

• Do I need massive scale / distributed architecture?
  → Milvus

• Do I need advanced ranking and retrieval logic?
  → Vespa

---

My default pick

If you haven’t decided yet: Qdrant is usually the best starting point for a self-hosted vector search stack because it gives you a strong combination of:

• performance
• simplicity
• filtering
• production readiness
• easy self-hosting

---

If you want, I can also give you:

1. a feature-by-feature comparison table,
2. a recommended architecture for your scale,
3. or a Docker Compose stack for one of these options.