Home / AI & GenAI / Vector Databases

AI & GenAICheat sheetIntermediate

Advisory: choosing a vector store - pgvector, Azure AI Search, OpenSearch, Qdrant, Milvus or a managed vector database

The vector store affects retrieval quality, operations effort and cost for years. Decide by scale, the need for hybrid keyword search and filters, where your data already lives, security and networking requirements, and who will operate it. Start simple and move only when evidence says you've outgrown it.

Deepak Bhardwaj2 min read

Applies toPostgreSQL with pgvector, Azure AI Search, Amazon OpenSearch / Elasticsearch, Qdrant, Milvus/Zilliz, Weaviate, Pinecone, MongoDB Atlas Vector Search, Databricks Vector Search, Cosmos DB vector search

Categories

Category Examples Strength
Vector in your existing database pgvector, MongoDB Atlas, Cosmos DB, Oracle AI Vector Search One system, transactions, joins with business data
Search engines with vectors Azure AI Search, OpenSearch, Elasticsearch Mature keyword + vector hybrid, filters, facets, analysers
Dedicated vector databases Qdrant, Milvus, Weaviate, Pinecone Vector-first features, scale, quantization options
Platform-native Databricks Vector Search, Bedrock Knowledge Bases backends Integrated sync with the data platform

Decision questions

  1. Scale: thousands, millions or billions of vectors? Growth rate?
  2. Hybrid search: do users search for codes, names, exact terms? → Strong keyword support matters.
  3. Filters: metadata and access control filters on every query?
  4. Where is the data? Keeping vectors next to source data simplifies sync and permissions.
  5. Operations: managed service or self-hosted (see the Qdrant standalone and vector HA posts)?
  6. Networking and security: private endpoints, customer-managed keys, regions.
  7. Cost model: per node, per capacity unit, per query, storage?

Typical fits

Situation Often a good fit
App already on PostgreSQL, up to tens of millions of vectors pgvector (see the consolidation transformation post)
Enterprise document search on Azure with hybrid + semantic ranking Azure AI Search
AWS, log/document search with vectors OpenSearch
Very large vector-heavy workloads, advanced quantization Qdrant / Milvus (self-hosted or cloud)
Data and pipelines on Databricks Databricks Vector Search

Evaluate with your data

  • Index a representative sample.
  • Measure recall@k against exact search, p95 latency at target QPS, filter performance, index build time, memory/cost.
  • Test ingestion updates/deletes and backup/restore (see the vector index DR post).

Avoid

  • Picking a dedicated vector DB for 100k vectors when your existing database can do it.
  • Running two vector stores for the same corpus without a reason.
  • Ignoring hybrid search; pure vector search struggles with exact identifiers.

Tip

Keep embeddings and chunk text in a durable source table (lake or database) regardless of the store. Rebuilding or switching stores then becomes a re-index job, not a re-embedding project.

Checklist

  • Scale, hybrid and filter needs documented
  • Data location and operations model considered
  • Security and networking requirements checked
  • Candidates tested on recall, latency, cost
  • Durable source of chunks/embeddings kept

Related fixes