Loading page…
pgvector Development
Vector similarity search inside PostgreSQL, our default when Postgres is already in the system and embeddings need to sit beside relational data.
Start Your Project
What pgvector is
pgvector is an open-source PostgreSQL extension that adds a vector column type and similarity operators. You store embeddings in an ordinary table and query them with SQL, using exact search or approximate indexes (HNSW or IVFFlat) for speed.
Its strength is that vectors are just another column. A query can combine similarity with ordinary filters and joins, inside one transaction, protected by the same backups, roles and row-level security as the rest of your data.
Suitability
And when it is not. Choosing the wrong tool costs more than any implementation detail.
Good fit for
- Applications already on PostgreSQL that need semantic search or RAG retrieval
- Retrieval that must be filtered by relational data, such as tenant, permissions, status or date
- Teams that want one database to operate, back up and secure rather than two
- Corpora in the range a single well-sized Postgres instance handles comfortably
Consider alternatives for
- Very large embedding collections with heavy query volume, where a purpose-built vector engine scales more predictably
- Workloads needing advanced vector-native features such as built-in hybrid ranking pipelines out of the box
- Teams without Postgres expertise who would be adopting Postgres only for this feature
In practice
01
RAG on an existing Postgres system
Adding retrieval to an application without introducing a new datastore, keeping documents, embeddings and permissions in one place.
02
Permission-aware retrieval
Combining similarity search with row-level security so a user can only retrieve what they are allowed to see.
03
Hybrid keyword and vector search
Pairing PostgreSQL full-text search with vector similarity in one query for results that respect both exact terms and meaning.
Considerations
Practical notes from running this in production.
- 01
Approximate indexes trade recall for speed. Tune index parameters against a labelled sample of real queries instead of accepting defaults, and measure recall rather than assuming it.
- 02
Index builds and memory use grow with the collection. Plan instance memory for the index, and rebuild on a copy when you change parameters.
- 03
Filtered approximate search can return fewer results than requested when a filter is very selective. Test selective filters explicitly, since this is the most common surprise in production.
- 04
Vectors are large. Storage, backup size and replication cost grow with embedding dimension, so choose the embedding model with that cost in mind.
FAQ
Is pgvector good enough, or do we need a dedicated vector database?
For many applications, pgvector is enough, and it avoids running and securing a second system. A dedicated store such as Qdrant, Weaviate or Pinecone earns its place when the collection is very large, query volume is high, or you need vector-specific features. We would rather start with pgvector and move if measurement shows a real limit than adopt a new database on speculation.
How does pgvector compare with Elasticsearch for search?
They solve different problems. Elasticsearch is strongest at keyword relevance, faceting and large-scale text analytics. pgvector is strongest at similarity over embeddings next to relational data. Some systems use both; many need only one. We decide from the actual queries users will run.
Do you use pgvector in production?
PostgreSQL is our primary relational database, and pgvector is the option we reach for first when a Postgres-based system needs retrieval. We do not publish benchmark numbers; sizing depends on your data and embedding model, and we test against your own sample.
How long does a pgvector development project take?
It depends on scope. A focused first release is usually weeks rather than months, while larger platforms are delivered in phases. We give a written estimate after a short discovery conversation and agree milestones before work starts.
How is pgvector development work priced?
We work either to a fixed scope with agreed milestones or on a time-and-materials basis, whichever suits how well the requirements are known. You receive the estimate and its assumptions in writing before committing.
By industry
Where we apply pgvector. Tiles marked "Guide" open a dedicated write-up; the rest open the industry page.
Related
- ServicesAI Development
- ServicesGenerative AI
- ServicesDatabase Development
- ServicesLLM Development
- SolutionsBusiness Automation
- SolutionsSaaS Platform Development
- IndustriesLegal Services
- IndustriesHealthcare
- IndustriesFinTech
TechnologiesPostgreSQL
TechnologiesPinecone
TechnologiesQdrant
TechnologiesWeaviate
TechnologiesLangChain
TechnologiesSupabase
Let’s talk about your pgvector project
Tell us what you are trying to build or fix. We will come back with scope, approach and an honest view on cost and timeline.