====== Supabase Vector ======

**Supabase Vector** leverages PostgreSQL with the [[pgvector|pgvector]] extension to enable efficient storage, indexing, and similarity search of high-dimensional vector embeddings within the Supabase platform. It combines vector search with Supabase's authentication, row-level security, Edge Functions, and real-time capabilities.((source [[https://supabase.com/features/vector-database|Supabase Vector Database]]))

===== pgvector Integration =====

Supabase integrates pgvector as a PostgreSQL extension, allowing vectors to be stored alongside relational data in the same schema:((source [[https://sparkco.ai/blog/mastering-supabase-vector-storage-a-2025-deep-dive|Supabase Vector Storage Deep Dive - Sparkco]]))

  * Enable with ''CREATE EXTENSION IF NOT EXISTS vector WITH SCHEMA extensions;''
  * Store embeddings in dedicated ''VECTOR'' columns
  * Support for HNSW and IVFFlat indexing
  * Hybrid semantic and keyword queries by combining embeddings with text columns
  * Standard SQL operations (joins, filters, aggregations) alongside vector search

===== Edge Functions =====

Supabase Edge Functions (serverless TypeScript/Deno functions) handle embedding generation and processing:((source [[https://supabase.com/docs/guides/ai|Supabase AI Documentation]]))

  * Process incoming data and generate embeddings in real-time
  * Integrate with OpenAI, Hugging Face, Cohere, and other embedding APIs
  * Upsert embeddings into pgvector tables
  * Deployable as database webhooks triggered on row changes
  * No external infrastructure required

===== Built-in Embedding Model =====

Since Supabase Edge Runtime v1.36.0, the **gte-small** model (384 dimensions) runs natively within Edge Functions:((source [[https://supabase.com/docs/guides/functions/examples/semantic-search|Supabase Semantic Search Example]]))

  * No external API calls required for embedding generation
  * ''const model = new Supabase.ai.Session('gte-small')''
  * Suitable for lightweight semantic search applications
  * Reduces latency and external dependencies

===== Similarity Search =====

Supabase exposes vector search through RPC functions callable via client SDKs:((source [[https://supabase.com/features/vector-database|Supabase Vector Database]]))

  * Create PostgreSQL functions using pgvector operators (''<=>'', ''<->'', ''<#>'')
  * Call via Supabase JS/Python SDKs using ''supabase.rpc()''
  * Support for top-k retrieval with distance metrics
  * Combinable with metadata filters and row-level security policies

**Vector Buckets** (launched 2026) provide S3-backed vector storage for larger-scale workloads:((source [[https://supabase.com/blog/vector-buckets|Supabase Vector Buckets]]))

  * Up to 50 million vectors per index
  * Queryable via JavaScript SDK and Postgres Foreign Data Wrappers
  * Multi-index buckets for multi-tenant applications

===== RAG Patterns =====

Supabase facilitates retrieval-augmented generation workflows:((source [[https://www.echoalgoridata.no/en/blog/rag-systems-supabase-guide|Building RAG Systems with Supabase - Echo Algori Data]]))

  * Store document chunks with their embeddings in pgvector tables
  * Retrieve top-k similar chunks via RPC similarity search
  * Feed retrieved context to LLMs (OpenAI, Anthropic, etc.) for generation
  * Hybrid RAG combining semantic search with keyword filtering
  * Edge Functions orchestrate the full embed-search-generate pipeline

===== Row-Level Security =====

Row-level security (RLS) applies to vector tables like any PostgreSQL relation:((source [[https://supabase.com/features/vector-database|Supabase Vector Database]]))

  * Enforce policies on SELECT, INSERT, UPDATE operations involving embeddings
  * Filter similarity searches by user ID or organization
  * GDPR-compliant data isolation for multi-tenant applications
  * No custom application-level access control needed

===== Automatic Embeddings Pipeline =====

Supabase provides an automated embedding generation pipeline:((source [[https://supabase.com/docs/guides/ai/automatic-embeddings|Supabase Automatic Embeddings]]))

  * **Triggers** -- detect content changes and enqueue embedding requests
  * **pgmq** -- queue embedding generation for processing and retries
  * **pg_net** -- asynchronous HTTP requests to Edge Functions from Postgres
  * **pg_cron** -- scheduled processing and retry of failed embeddings
  * Generic and reusable across multiple tables and content types

===== Supabase Vector vs Alternatives =====

^ Aspect ^ Supabase Vector ^ Standalone pgvector ^ Pinecone ^
| Setup | Instant with Supabase project | Self-managed PostgreSQL | Managed SaaS |
| Auth/RLS | Built-in with Supabase Auth | Manual implementation | API key-based |
| Scale | Millions (pgvector) + 50M (Vector Buckets) | Depends on PostgreSQL tuning | Billions with auto-sharding |
| Cost | Included in Supabase pricing | Free extension | Usage-based subscription |
| Best For | AI apps needing auth, real-time, and relational data | Custom PostgreSQL deployments | Pure vector workloads at scale |

Supabase Vector excels in applications requiring unified relational and vector data with built-in authentication and security.((source [[https://sparkco.ai/blog/mastering-supabase-vector-storage-a-2025-deep-dive|Supabase Vector Storage Deep Dive - Sparkco]]))

===== See Also =====

  * [[pgvector|pgvector]]
  * [[pinecone|Pinecone]]
  * [[hugging_face|Hugging Face]]

===== References =====