@supavec/supabase-ai

A TypeScript SDK for building RAG (Retrieval-Augmented Generation) applications with Supabase and pgvector.

Features

• 🔍 Semantic Search: Powerful vector similarity search with pgvector
• 🤖 OpenAI Integration: Seamless OpenAI embeddings integration
• 📦 Type-Safe: Full TypeScript support with comprehensive type definitions
• 🚀 Easy Integration: Works with your existing Supabase client
• 🔧 Flexible Configuration: Customizable similarity thresholds and more
• 📊 Metadata Filtering: Advanced filtering with JSON metadata support

Installation

npm install @supavec/supabase-ai

Prerequisites

Before using this SDK, you need:

1. Supabase project with pgvector extension enabled
2. Database tables set up for storing embeddings
3. RPC functions for similarity search

Required Database Setup

-- Enable pgvector extension
CREATE EXTENSION IF NOT EXISTS vector;

-- Create documents table
CREATE TABLE documents (
  id uuid PRIMARY KEY DEFAULT gen_random_uuid(),
  content text NOT NULL,
  embedding vector(1536), -- Adjust dimensions based on your model
  metadata jsonb DEFAULT '{}',
  user_id uuid,
  created_at timestamptz DEFAULT now()
);

-- Create index for vector similarity search
CREATE INDEX ON documents USING ivfflat (embedding vector_cosine_ops);

-- Create RPC function for similarity search
CREATE OR REPLACE FUNCTION match_documents(
  query_embedding vector(1536),
  match_threshold float,
  match_count int,
  table_name text,
  filters jsonb DEFAULT '{}'
)
RETURNS TABLE (
  id uuid,
  content text,
  metadata jsonb,
  similarity float
)
LANGUAGE plpgsql
AS $$
BEGIN
  RETURN QUERY
  EXECUTE format('
    SELECT 
      id,
      content,
      metadata,
      1 - (embedding <=> $1) as similarity
    FROM %I
    WHERE 1 - (embedding <=> $1) > $2
    ORDER BY embedding <=> $1
    LIMIT $3
  ', table_name)
  USING query_embedding, match_threshold, match_count;
END;
$$;

Quick Start

// ES modules
import { createClient } from '@supabase/supabase-js'
import { SupabaseAI } from '@supavec/supabase-ai'

// CommonJS
// const { createClient } = require('@supabase/supabase-js')
// const { SupabaseAI } = require('@supavec/supabase-ai')

// Initialize Supabase client
const supabase = createClient(
  process.env.SUPABASE_URL!,
  process.env.SUPABASE_ANON_KEY!
)

// Initialize SupabaseAI
const ai = new SupabaseAI(supabase, {
  apiKey: process.env.OPENAI_API_KEY!,
  embeddings: {
    provider: 'openai', // Currently only OpenAI is supported
    model: 'text-embedding-3-small',
    table: 'documents',
    threshold: 0.8
  }
})

// Store documents (pre-chunked)
await ai.embeddings.store([
  {
    content: 'The quick brown fox jumps over the lazy dog.',
    metadata: { title: 'Example Document', type: 'text' }
  }
]) // Uses default table from config

// Search documents
const results = await ai.embeddings.search('fox jumping', {
  table: 'documents',
  limit: 5,
  threshold: 0.8
})

console.log(results)

API Reference

SupabaseAI

Main client class for interacting with the SDK.

const ai = new SupabaseAI(supabaseClient, options)

Options

• apiKey: string - OpenAI API key
• embeddings?: object - Embeddings configuration
  • provider?: "openai" - Embedding provider (currently only OpenAI is supported)
  • model?: string - OpenAI model name (default: 'text-embedding-3-small')
  • table?: string - Default table for operations
  • threshold?: number - Default similarity threshold (default: 0.8)

EmbeddingsClient

Handle embedding operations and semantic search.

Methods

store(data, options?)

Store documents with automatic embedding generation. Each item in the data array should represent a single, pre-chunked piece of content.

// Using default table from config
await ai.embeddings.store([
  {
    content: 'Document text content (pre-chunked)',
    metadata: { title: 'Document Title', category: 'tech' },
    user_id: 'user123'
  }
])

// Or specify options
await ai.embeddings.store([...], {
  table: 'custom_table',
  batchSize: 50
})

Store Options:

• table?: Table name (optional if embeddings.table is set)
• generateId?: Generate IDs using SDK (default: false, lets database auto-generate)
• batchSize?: Number of records to insert per batch (default: 100)

ID Handling:

• By default, the SDK lets your database auto-generate IDs (recommended for UUID primary keys)
• If your data includes id fields, they will be used
• Set generateId: true to force SDK-generated IDs when no ID is provided

LangChain Integration: The store method accepts both native format and LangChain Documents directly:

import { Document } from '@langchain/core/documents'

// LangChain Documents work directly
const langchainDocs = [
  new Document({ pageContent: "LangChain document content", metadata: { source: "web" } })
]
await ai.embeddings.store(langchainDocs)

// Mixed formats also work
await ai.embeddings.store([
  new Document({ pageContent: "LangChain doc", metadata: { type: "langchain" } }),
  { content: "Native format doc", metadata: { type: "native" } }
])

Note: Content should be pre-chunked using your preferred method (LangChain text splitters, etc.) before passing to the store method.

search(query, options)

Perform semantic search on stored documents.

const results = await ai.embeddings.search('search query', {
  table: 'documents',
  limit: 10,
  threshold: 0.8,
  filters: { user_id: 'user123' },
  metadata: { category: 'tech' },
  select: 'id, content, metadata, created_at',
  orderBy: 'similarity',
  includeDistance: true
})

Search Options:

• table?: Table name (optional if embeddings.table is set)
• limit?: Maximum results (default: 10)
• threshold?: Similarity threshold (default: 0.8)
• filters?: SQL-style filters for table columns
• metadata?: JSON metadata filters
• select?: Custom SELECT clause
• orderBy?: Sort order ('similarity' | 'created_at' | column name)
• includeDistance?: Include similarity scores in results
• rpc?: Custom RPC function name

create(input, options?)

Generate embeddings for text input.

const embeddings = await ai.embeddings.create(['text1', 'text2'])

similarity(text1, text2)

Calculate similarity between two text strings.

const score = await ai.embeddings.similarity('hello world', 'hello there')

Advanced Usage

LangChain Integration

Complete example using LangChain text splitters:

import { RecursiveCharacterTextSplitter } from '@langchain/textsplitters'
import { Document } from '@langchain/core/documents'

// Split large document using LangChain
const textSplitter = new RecursiveCharacterTextSplitter({
  chunkSize: 1000,
  chunkOverlap: 200,
})

const docs = await textSplitter.createDocuments([
  'Your large document text here...',
], [{ source: 'document.pdf', author: 'John Doe' }])

// Store directly - no conversion needed
await ai.embeddings.store(docs)

// Search works the same way
const results = await ai.embeddings.search('query about the document')

Complex Search Queries

const results = await ai.embeddings.search('machine learning concepts', {
  table: 'knowledge_base',
  limit: 20,
  threshold: 0.85,
  filters: {
    created_at: { gte: '2024-01-01' },
    status: 'published',
    user_id: userId
  },
  metadata: {
    category: 'technical',
    tags: ['ai', 'ml']
  },
  select: 'id, content, title, metadata, created_at, author',
  orderBy: 'similarity',
  includeDistance: true
})

Batch Operations

// Pre-chunk your documents using your preferred method
const documents = [
  { content: 'Document 1 chunk 1', metadata: { type: 'article', chunk: 1 } },
  { content: 'Document 1 chunk 2', metadata: { type: 'article', chunk: 2 } },
  { content: 'Document 2 chunk 1', metadata: { type: 'blog', chunk: 1 } },
  // ... more pre-chunked documents
]

// With custom IDs
const documentsWithIds = [
  { id: 'doc1-chunk1', content: 'Document 1 chunk 1', metadata: { type: 'article' } },
  { id: 'doc1-chunk2', content: 'Document 1 chunk 2', metadata: { type: 'article' } },
]

// Using default table from config
await ai.embeddings.store(documents, {
  batchSize: 50 // Process in batches of 50
})

// Or specify custom table
await ai.embeddings.store(documents, {
  table: 'custom_documents',
  batchSize: 50
})

Error Handling

The SDK provides comprehensive error handling:

import { 
  SupabaseAIError, 
  EmbeddingProviderError, 
  DatabaseError, 
  ValidationError 
} from '@supavec/supabase-ai'

try {
  await ai.embeddings.search('query', { table: 'documents' })
} catch (error) {
  if (error instanceof ValidationError) {
    console.error('Validation error:', error.message)
  } else if (error instanceof DatabaseError) {
    console.error('Database error:', error.message)
  } else if (error instanceof EmbeddingProviderError) {
    console.error('Provider error:', error.message)
  }
}

TypeScript Support

Full TypeScript support with comprehensive type definitions:

import { SearchResult, SearchOptions, StoreData } from '@supavec/supabase-ai'

const searchOptions: SearchOptions = {
  table: 'documents',
  limit: 10,
  threshold: 0.8
}

const results: SearchResult[] = await ai.embeddings.search('query', searchOptions)

Contributing

Contributions are welcome! Please read our contributing guidelines and submit pull requests.

Development

• Run tests: npm test
• Build project: npm run build
• All PRs are automatically tested via GitHub Actions CI

License

MIT License - see LICENSE file for details.