Vector database cost comparison
Choosing where to store your embeddings is partly a feature decision and partly a bill. This tool estimates the monthly cost of the same workload across four popular options — Pinecone, Weaviate Cloud, Qdrant Cloud, and self-hosted pgvector — from three inputs: how many vectors you store, their dimensionality, and how many queries you run per month.
How it works
First the tool computes raw storage as vectors × dimensions × 4 bytes for
float32 embeddings, then adds about 30% for index and metadata overhead. Each
provider applies its own model on top of that storage figure: a managed serverless
service charges per gigabyte stored plus a per-query rate; pod-based pricing maps
the storage tier to an instance fee; and pgvector is modelled as a flat managed
Postgres instance whose cost barely moves with scale until you outgrow the box.
The side-by-side table then ranks the monthly totals so you can see where the
break-even points fall as your data and traffic grow.
Storage size in practice
A typical RAG application embeds documents using OpenAI’s text-embedding-3-small,
which produces 1,536-dimensional vectors. For illustration: storing 500,000 vectors
at 1,536 dimensions requires 500,000 × 1,536 × 4 bytes ≈ 3.1 GB of raw float32
storage before index overhead. Adding 30% overhead brings the total to roughly 4 GB.
Switching to a 768-dimension model (such as a compressed sentence-transformer) nearly
halves that footprint, which matters both for storage cost and for in-memory index performance.
When to choose each option
Pinecone Serverless suits applications with unpredictable or bursty traffic where you want zero infrastructure management. The pay-per-query model is economical at low traffic but can become expensive at sustained high QPS.
Weaviate Cloud combines vector search with structured object storage and GraphQL querying. It is a good fit for use cases where you want to combine semantic search with traditional filters on rich metadata without running a separate database.
Qdrant Cloud is designed for high-recall, high-performance similarity search and offers strong filtering on payload metadata. Its pod-based pricing is more predictable at high QPS than per-query models.
pgvector (self-hosted or managed Postgres) is the lowest-cost option for small to mid-scale use cases and teams already running Postgres. The trade-off is that you are responsible for index tuning (HNSW parameters, maintenance), scaling, and recall quality — there is no managed index optimizer working on your behalf.
Key cost drivers to watch
- Dimensions: the single largest lever. Halving dimensions roughly halves storage cost and speeds up both indexing and retrieval.
- Index type: HNSW indexes give high recall with fast queries but consume significant memory; IVF-Flat is cheaper to store but slower for large datasets.
- Replicas: managed services charge per replica. Adding a second replica for high availability typically doubles the storage bill.
- Query volume: serverless models charge per read; pod-based models charge for the instance whether it is busy or idle.
Tips and notes
- Estimates are local and editable. Refresh the inputs whenever provider list prices change — managed-cloud pricing shifts frequently.
- Free tiers exist. All four options have a free or trial tier suitable for prototyping; the comparison becomes meaningful once you project production scale.
- Recall matters. A cheaper option that misses 10% of the best results is not actually cheaper if it degrades the quality of your application.