Vectorizer API reference

Automate embedding generation and synchronization for your PostgreSQL data

A vectorizer provides a powerful and automated way to generate and manage LLM embeddings for your Postgres data, keeping them synchronized with your source data automatically.

What is a vectorizer?

A vectorizer automates the entire embedding workflow:

• Automated embedding generation: Create embeddings for table data automatically
• Automatic synchronization: Triggers keep embeddings in sync with source data
• Background processing: Async processing minimizes impact on database operations
• Scalability: Batch processing handles large datasets efficiently
• Highly configurable: Customize embedding models, chunking, formatting, indexing, and scheduling

Key features

• Multiple AI providers: OpenAI, Ollama, Cohere, Voyage AI, and LiteLLM support
• Efficient storage: Separate tables with appropriate indexing for similarity searches
• View creation: Automatic views join source data with embeddings
• Access control: Fine-grained permissions for vectorizer objects
• Monitoring: Built-in tools to track queue status and performance

Quick start

Create a basic vectorizer

SELECT ai.create_vectorizer(
    'blog.posts'::regclass,
    embedding => ai.embedding_openai('text-embedding-3-small', 768),
    chunking => ai.chunking_character_text_splitter(512)
);

Table destination (separate embeddings table)

SELECT ai.create_vectorizer(
    'website.blog'::regclass,
    destination => ai.destination_table(
        target_table => 'blog_embeddings_store',
        view_name => 'blog_embeddings'
    ),
    loading => ai.loading_column('content'),
    embedding => ai.embedding_ollama('nomic-embed-text', 768),
    chunking => ai.chunking_character_text_splitter(128, 10)
);

Column destination (embedding in source table)

SELECT ai.create_vectorizer(
    'products'::regclass,
    destination => ai.destination_column('description_embedding'),
    loading => ai.loading_column('description'),
    embedding => ai.embedding_openai('text-embedding-3-small', 768),
    chunking => ai.chunking_none()  -- Required for column destination
);

Configuration functions

Core functions

• create_vectorizer(): create and configure a new vectorizer
• drop_vectorizer(): remove a vectorizer and clean up resources

Destination configuration

• destination_table(): store embeddings in a separate table (default)
• destination_column(): store embeddings in the source table

Loading configuration

• loading_column(): load data from a column
• loading_uri(): load data from a file URI

Parsing configuration

• parsing_auto(): auto-detect document format (default)
• parsing_none(): no parsing for text data
• parsing_docling(): parse documents with Docling
• parsing_pymupdf(): parse PDFs with PyMuPDF

Chunking configuration

• chunking_character_text_splitter(): split by character count
• chunking_recursive_character_text_splitter(): recursive splitting (default)

Embedding configuration

• embedding_openai(): OpenAI embedding models
• embedding_ollama(): local Ollama models
• embedding_litellm(): unified API for 100+ providers
• embedding_voyageai(): Voyage AI models

Formatting configuration

• formatting_python_template(): format with Python templates

Indexing configuration

• indexing_default(): default HNSW indexing
• indexing_diskann(): DiskANN indexing
• indexing_hnsw(): HNSW indexing with options
• indexing_none(): no automatic indexing

Scheduling configuration

• scheduling_default(): run every 5 minutes
• scheduling_timescaledb(): use TimescaleDB job scheduling
• scheduling_none(): disable automatic scheduling

Processing configuration

• processing_default(): default processing settings

Access control

• grant_to(): specify user permissions

Management functions

• enable_vectorizer_schedule(): resume automatic processing
• disable_vectorizer_schedule(): pause automatic processing

Monitoring

• vectorizer_status: view vectorizer status and statistics
• vectorizer_queue_pending(): check pending work items