Supabase Vector Database & pgvector: Complete Guide 2026

TL;DR Supabase vector search uses the pgvector PostgreSQL extension to store and query high-dimensional embeddings directly inside your existing relational database. Paired with Claude or OpenAI embeddings, it delivers production-grade semantic search without forcing you onto a dedicated vector vendor. For most teams shipping in 2026, Supabase + pgvector is cheaper, simpler, and less risky than Pinecone, Weaviate, or Qdrant as standalone services.

Updated June 2026 by Marianella Saavedra Terkes. 10 minute read. Kreante is a Bubble Gold Partner, Webflow Expert, FlutterFlow certified agency, and Anthropic Claude Partner Network member. 265+ projects, 110+ clients across 35+ countries, rated 4.9/5 on Clutch.

What is Supabase Vector?

Supabase Vector is the vector storage and similarity-search layer built into Supabase, powered by the open-source pgvector extension for PostgreSQL. It lets you store float arrays (embeddings) as a native column type and run approximate nearest-neighbor (ANN) searches inside the same database where your users, content, and transactions already live.

No separate vector service. No extra API key. No duplicated data pipeline.

Enable it with one SQL command:

create extension if not exists vector;

Then add a vector column to any table:

alter table documents add column embedding vector(1536);

That is the entire infrastructure change.

pgvector: What It Actually Does

pgvector adds three things to PostgreSQL:

1. A native vector column type supporting up to 16,000 dimensions (pgvector 0.7+, released Q1 2025).
2. Two index types: HNSW (Hierarchical Navigable Small World) for fast approximate search, and IVFFlat for memory-efficient large datasets.
3. Three distance operators: cosine distance, L2/Euclidean, and inner product.

HNSW is the default recommendation for 2026. It builds a graph structure that keeps query latency under 10ms at the 99th percentile for datasets up to ~5 million vectors on a standard Supabase Pro instance.

Supabase pgvector vs Pinecone vs Weaviate vs Qdrant in 2026

Dimension | Supabase + pgvector | Pinecone | Weaviate | Qdrant

Hosting model | Managed Postgres (or self-host) | Fully managed SaaS | Managed or self-host | Managed or self-host

Vendor lock-in | Low (standard SQL) | High (proprietary API) | Medium | Medium

Max dimensions | 16,000 (pgvector 0.7+) | 20,000 | 65,535 | 65,535

Hybrid search (vector + SQL filter) | Native JOIN support | Limited metadata filtering | GraphQL-based | Payload filtering

Free tier (2026) | 500 MB storage, 2 projects | 5 indexes, 100k vectors | 1 sandbox cluster | 1 free cluster

Pro pricing | $25/month (includes full Postgres) | $70/month (1M vectors) | ~$25/month+ | ~$25/month+

Latency (1M vectors, HNSW) | 10-30ms | 5-15ms | 10-25ms | 8-20ms

Open source | Yes (pgvector + Supabase) | No | Yes | Yes

Claude/OpenAI embeddings | Direct via Edge Functions | Yes | Yes | Yes

Existing relational data joins | Native | No | No | No

The headline number: Pinecone at $70/month for 1M vectors versus Supabase Pro at $25/month where vector search is included in the same instance running your auth, storage, and relational tables. That math gets more extreme as you scale.

Why Vendor Lock-in Matters in 2026

Gartner's 2026 AI Infrastructure report flags vector database fragmentation as a top-5 risk for AI-native product teams. The specific concern: migrating embeddings between proprietary vector services requires re-indexing every document, which at 10M vectors can cost thousands in embedding API calls alone.

With Supabase, your data lives in standard PostgreSQL. You can dump it, move it, or spin up a self-hosted instance without rewriting your query logic.

How Kreante Builds With Supabase Vector

One client in Kreante's portfolio (a Latin American edtech platform) needed semantic document search across 40,000 course resources in Spanish and Portuguese. The requirements: sub-200ms search, filtering by subject and language, and no third-party vector service that would complicate their data residency policy.

The stack Kreante shipped:

• Supabase Pro for PostgreSQL + pgvector storage
• Claude via the Anthropic API for embedding generation
• HNSW index with cosine distance
• Supabase Edge Functions to handle embedding creation on document upload
• n8n for the ingestion pipeline (PDF to chunk to embed to insert)

Result: average query latency of 45ms at 40,000 vectors, zero additional infrastructure cost beyond the $25/month Supabase Pro plan, and all data stays inside their existing database.

This is the pattern Kreante repeats across projects: keep the stack unified, avoid paying twice for infrastructure that should talk to itself.

Want to see if this applies to your product? Book a 30-minute call with the Kreante team.

Claude Embeddings + Supabase: The 2026 Setup

As an Anthropic Claude Partner Network member since April 2026, Kreante uses Claude's embedding models as the default for new projects. The workflow:

1. Chunk your documents (recommended: 512 tokens with 50-token overlap).
2. Send chunks to the Claude embedding endpoint.
3. Insert the returned vector into your Supabase vector column alongside the source text and metadata.
4. At query time, embed the user's search string and run the ANN query.

Claude's embeddings handle multilingual content well without separate models per language. For a platform serving Spanish, Portuguese, and English users, that alone removes one infrastructure layer.

2026 Pricing: Supabase Vector

Plan | Monthly cost | Vector storage | Projects

Free | $0 | 500 MB total DB (vectors included) | 2

Pro | $25 | 8 GB total DB | Unlimited

Team | $599 | Custom | Custom

Enterprise | Custom | Custom | Custom

pgvector is included on all plans at no extra cost.

When NOT to Use Supabase Vector

Supabase + pgvector is not the right choice in every case:

• Datasets above 50M vectors on a single node: PostgreSQL's HNSW performance degrades at extreme scale.
• Sub-5ms latency requirements: Pinecone's in-memory indexes consistently outperform pgvector at tail latency percentiles.
• Teams already deep in the Weaviate or Qdrant ecosystem: migration cost may not justify switching.
• Pure vector workloads with no relational data: a specialized vector service may be simpler to operate.
• Compliance environments blocking managed cloud databases: self-hosted pgvector is an option, but then you own the operations burden.

External Benchmark Reference

The ann-benchmarks project (ann-benchmarks.com), updated Q1 2026, shows HNSW in pgvector achieving 95%+ recall at 10-querier parallelism for 1M-vector datasets using cosine distance. Pinecone's managed HNSW achieves slightly lower latency at equivalent recall, primarily because of its dedicated in-memory architecture. The gap at 1M vectors is roughly 2x in raw QPS, but most web application workloads never hit the concurrency levels where that gap matters.

The practical conclusion: if your p99 query volume is under 50 requests per second, Supabase pgvector's performance is indistinguishable from dedicated vector services for end users.

FAQ

What is pgvector and how does it work with Supabase?
pgvector is an open-source PostgreSQL extension that adds a vector column type and approximate nearest-neighbor search operators. Supabase includes pgvector on all plans, so you can store embeddings and run semantic search inside your existing Postgres database without a separate vector service.

How many vectors can Supabase pgvector handle in 2026?
HNSW indexes in pgvector perform well up to around 5-10 million vectors on a Supabase Pro instance before you need to consider partitioning or upgrading to a larger compute tier. For most production applications, this is more than sufficient.

Is Supabase Vector cheaper than Pinecone?
For the majority of use cases, yes. Supabase Pro at $25/month includes your entire backend, and pgvector storage is billed at standard Postgres storage rates. Pinecone charges separately per vector and per read operation, which adds up quickly above 500,000 vectors or high query frequency.

Can I use Claude embeddings with Supabase?
Yes. You generate embeddings via the Anthropic API (or any embedding provider) and insert the resulting float array into your Supabase vector column. There is no coupling between the embedding model and the database. Kreante is an Anthropic Claude Partner Network member and uses this setup on client projects.

What is the difference between HNSW and IVFFlat indexes in pgvector?
HNSW builds a navigable graph and provides faster query times with no training step. IVFFlat clusters vectors into lists and requires a training phase on existing data, but uses less memory. For 2026 projects starting fresh, HNSW is the default recommendation.

Does Supabase support hybrid search (vector + keyword)?
Yes. Because pgvector runs inside PostgreSQL, you can combine a vector similarity sort with any standard SQL filter: WHERE clauses, JOINs, full-text search using tsvector, or any combination.

What dimensions should I use for embeddings in 2026?
OpenAI text-embedding-3-small outputs 1536 dimensions by default (reducible to 512). For most document search applications, 512-1536 dimensions balances accuracy and performance.

How does Supabase Vector avoid vendor lock-in?
Your embeddings and metadata are stored in standard PostgreSQL tables. You can export the data with a standard pg_dump, run the same queries on any PostgreSQL instance with pgvector installed, or migrate to self-hosted infrastructure. No proprietary API format, no re-indexing required.

Bottom Line

Supabase + pgvector is the pragmatic choice for semantic search in 2026 if you already use Postgres, care about data ownership, or want one bill instead of four. It is not the highest-throughput option at extreme scale, but for the application layer that most product teams actually build, it is fast enough, cheap enough, and simple enough to be the default.

Dedicated vector services like Pinecone make sense when you have tens of millions of vectors and millisecond-level latency SLAs. For everything else, keeping search inside your existing database is the right engineering call.

If you want a second opinion on your vector search architecture or a team to build it, schedule a free 30-minute consultation with Kreante.

Related Articles

• Supabase: The Best Open Source Alternative to Firebase for Developers
• Supabase vs Xano: Choosing the Right Backend for Your Project
• Firebase: A Top No-Code Tool to Streamline Your App Development Process