pgvector vs Qdrant: when vectors should stay in PostgreSQL
An even-handed look at pgvector on PostgreSQL and Qdrant, with sources, for teams building RAG or search who already hold data in PostgreSQL.
45 minutes with an engineer: we review your workload and sketch the design.
- ✓OSStend runs pgvector and Qdrant
- ✓Sources checked 15 Nov 2026
Choose pgvector when your data already lives in PostgreSQL and you want one database, one backup and SQL joins with your vectors. Choose Qdrant when vector search is its own service, at a scale or filter load where a dedicated engine pays off. Both are open source: pgvector 0.8.7 (1 Oct 2026) and Qdrant 1.20.1 (13 Nov 2026).
- PostgreSQL
- MySQL
- MongoDB
- Redis
- Valkey
- OpenSearch
- Kafka
- RabbitMQ
- AWS
- Google Cloud
- Azure
- Kubernetes
- OpenShift
- Prometheus
- Grafana
- Datadog
- Dynatrace
- Splunk
pgvector and Qdrant at a glance
pgvector is an extension that adds vector search to PostgreSQL; Qdrant is a standalone vector database written in Rust. The table summarises what each project's own README and documentation state, read on 15 Nov 2026.
| Factor | pgvector on PostgreSQL | Qdrant |
|---|---|---|
| What it is | Open-source vector similarity search for Postgres, as an extension (README) | A vector similarity search engine and vector database that stores points, which are vectors with an additional payload (README) |
| Licence and current version | 0.8.7, released 1 Oct 2026 (changelog); permissive PostgreSQL-style terms (LICENSE) | 1.20.1, published 13 Nov 2026; Apache 2.0 (GitHub, releases) |
| Index types | Exact search, HNSW and IVFFlat; indexes cover up to 2,000 dimensions for vector, 4,000 for halfvec and 64,000 for bit (README) | HNSW as the dense vector index, an exact sparse vector index, and payload indexes for filtering (indexing docs) |
| Filtering | With approximate indexes, filters apply after the index scan; iterative scans (0.8.0 and later), partial indexes and partitioning help (README) | A filterable HNSW index extends the graph with edges based on indexed payload values; create payload indexes before ingesting data (indexing docs) |
| Hybrid search | Postgres full-text search combined with vector results, merged with Reciprocal Rank Fusion or a cross-encoder (README) | Query API with prefetch (from v1.10.0): dense, sparse and BM25 full-text queries, fused with RRF or DBSF (hybrid queries docs) |
| Where it runs | Wherever PostgreSQL 13 or later runs: your cloud account, a data center or a provider | Self-hosted, or Qdrant Cloud on AWS, Azure and GCP; the pricing page also lists Hybrid Cloud on your own infrastructure and Private Cloud (pricing) |
| Managed options | Any managed PostgreSQL service that offers the extension; OSStend runs it on managed PostgreSQL | Qdrant Cloud: a Free tier (0.5 vCPU, 1 GB RAM, 4 GB disk), usage-based Standard and a Premium tier with minimum spend (pricing) |
| With your relational data | ACID compliance, point-in-time recovery and JOINs with the rest of Postgres (README) | Points hold a vector plus payload; conditions on payload and point id are applied at search time (filtering docs) |
How we compared: each project's README, documentation and pricing page, read on 15 Nov 2026, with the link beside each fact. No benchmark was run and none is quoted. OSStend runs both PostgreSQL with pgvector and a dedicated Qdrant, so it has no stake in which one you pick; read the page knowing it sells both. Sources: pgvector README, Qdrant README, Qdrant docs and Qdrant pricing.
Where pgvector wins
pgvector wins when the vectors belong beside data that already lives in PostgreSQL. You keep one database, one backup and recovery story, and you filter and join with plain SQL.
- ✓One transaction. The pgvector README lists ACID compliance, JOINs and point-in-time recovery, and says replication works through the write-ahead log (README, checked 15 Nov 2026). The vector and the business row commit together.
- ✓One recovery story. Restore the database to a point in time and the vectors come back with the rest of the data. There is no second system to line up with it.
- ✓Existing access control. SQL joins and the permissions you already set in PostgreSQL apply to the vectors too.
- ✓One thing on call. A PostgreSQL team already runs backups, upgrades and monitoring; an extension adds index build time and memory to that work, not a new system.
Where Qdrant wins
Qdrant wins when vector search is its own service. Its documentation describes a dedicated engine with filtering built into the vector index, hybrid queries in one API, and horizontal scaling on its own.
- ✓Built for vector-first work. Qdrant describes itself as a vector similarity search engine and database, with sharding and replication to scale horizontally and resize collections with zero downtime (README, checked 15 Nov 2026).
- ✓Filtering inside the index. Its filterable HNSW index adds edges based on indexed payload values so that strict filters do not break the graph (indexing docs, checked 15 Nov 2026).
- ✓Hybrid queries in one API. Dense vectors, sparse vectors and BM25 full-text queries can be combined with prefetch and fused with RRF or DBSF (hybrid queries docs, checked 15 Nov 2026).
- ✓Scale the vector service separately. It runs as its own cluster, self-hosted or on Qdrant Cloud, so vector load does not share a server with your transactional database (pricing, checked 15 Nov 2026).
None of these is a reason to avoid pgvector in itself. They become decisive at a scale or filter load that you have measured on your own data.
Decision table: which one for your workload
If your data is already in PostgreSQL and the workload sits under the scale you have measured, start with pgvector. Move to a dedicated engine when a measured limit, not a guess, says so.
| If | Choose |
|---|---|
| Your source data is already in PostgreSQL and you have not yet hit a measured limit | pgvector: one database, one backup, SQL joins |
| Vector search is a separate service with no relational data beside it | Qdrant, or another dedicated vector engine (six compared in best open source vector databases) |
| Vectors must stay in your own cloud account or data center | Either can run there (pgvector on PostgreSQL, or self-hosted or Hybrid Cloud Qdrant); decide on the other rows |
| You have no capacity to run another database | pgvector, if PostgreSQL is already run for you; Qdrant Cloud if you want the vector service run by its vendor |
| Heavy filtered search at high query rates | Test both on your own vectors, filters and recall target |
Running it in production
pgvector adds index build time and memory to a PostgreSQL cluster you already operate. Qdrant adds a second cluster to back up, upgrade and monitor.
For pgvector, the README notes that HNSW indexes build more slowly and use more memory than IVFFlat, and warns when the graph no longer fits into maintenance_work_mem (README, checked 15 Nov 2026). Sizing and tuning are covered in pgvector in production. For Qdrant, the new work is a separate cluster with its own snapshots, upgrades and shard layout; the Qdrant Cloud tiers shift part of that work to the vendor. If PostgreSQL is already run for you, see managed PostgreSQL.
Cost model
pgvector costs whatever the PostgreSQL cluster it runs in costs. Qdrant costs self-hosted infrastructure, or Qdrant Cloud usage on top of the cluster you already run.
On its pricing page, Qdrant lists a Free tier with a single node (0.5 vCPU, 1 GB RAM, 4 GB disk), a usage-based Standard tier for production, and a Premium tier with a minimum spend; it also offers a pricing calculator (Qdrant pricing, checked 15 Nov 2026). pgvector has no price of its own: its cost is the extra memory, storage and replicas the vectors need in PostgreSQL. OSStend runs pgvector on managed PostgreSQL and a dedicated Qdrant, and quotes each per design; see pricing.
Where OSStend fits, and when it does not
OSStend fits whichever way this comparison lands: it runs pgvector on managed PostgreSQL and a dedicated Qdrant, in your own environment. It does not fit when you want a hosted vector API with no cluster to design, or a free tier.
Contact us if
- ✓Your source data is in PostgreSQL and the vectors belong beside it.
- ✓Embeddings must stay in your cloud account or data center.
- ✓You want vector search, backups, replicas and on-call under one operator.
Not a fit if
- ✕You want Qdrant Cloud, Pinecone or another vendor-run vector API with nothing to design; see the Qdrant Cloud comparison.
- ✕You want a hosted vector API with no database to operate.
For Pinecone as the dedicated option, see pgvector vs Pinecone. For the wider retrieval decision, see vector database for RAG and the compare hub.
FAQ
Is pgvector fast enough for production RAG?
It can be, when the index fits in memory, the filters are handled with iterative scans, partial indexes or partitioning, and you have tested recall and latency on your own vectors. This page quotes no benchmark; run your own workload before deciding.
Can I use pgvector and Qdrant together?
Yes. One split keeps vectors and business data in PostgreSQL and moves a vector-heavy, filter-heavy collection to Qdrant. The cost is a second system to run and a sync path between them, so do it only when a measured limit requires it.
Can OSStend run Qdrant?
Yes. OSStend runs a dedicated Qdrant in your environment, such as your cloud account or data center, designed for your workload, and runs pgvector on managed PostgreSQL too. See Qdrant Cloud pricing and alternatives.
Related pages
- GuideBest open source vector databases (2026)The best open source vector databases in 2026 compared: pgvector, Qdrant, Milvus, Weaviate and more, with the method we used and where each one fits.
- AI data · PostgreSQLpgvector in production: indexes and sizingpgvector in production: HNSW against IVFFlat, memory sizing, recall against latency, filtering, replicas and pgvectorscale, with links to the pgvector docs.
- PostgreSQLManaged PostgreSQL service and Postgres hostingManaged PostgreSQL service, compared: how to choose between RDS, Aiven, Crunchy Bridge, Neon and a tailored cluster run 24/7 in your cloud or ours.
- PricingDatabase as a service pricing: dedicated, tailoredDatabase as a service pricing from OSStend: dedicated, tailored clusters priced by engine, size, SLA tier and location. Enterprise only, with no free tier.
- AlternativeQdrant Cloud pricing and alternatives (2026)Qdrant Cloud pricing tiers and SLAs from its published page, and dedicated Qdrant run by OSStend engineers in your own environment, next to PostgreSQL.
- Comparisonpgvector vs Pinecone for enterprise workloadspgvector vs Pinecone for enterprise workloads: where each wins on operations, data location, scale, hybrid search and cost model. Sourced, 30 Oct 2026.
Talk to an engineer about pgvector
Bring the workload and where the data may live. An engineer will give a straight view on whether pgvector on PostgreSQL fits.
- ✓A design sketch for your workload, not a plan menu
- ✓A zero-downtime cutover path for what you run today
- ✓A price for that design, after the review