Skip to main content
Store embeddings directly in the source table as a new column. This approach requires a one-to-one relationship between source data and embeddings, so chunking must be disabled. Ideal when source text is short or chunking is done upstream.

​
Key features

  • Adds vector column directly to source table
  • No separate view created
  • Requires chunking_none() (no chunking)
  • Exactly one embedding per row
  • Simpler schema with fewer objects

​
Workflow

  1. Application inserts data with NULL in embedding column
  2. Vectorizer detects the NULL value
  3. Vectorizer generates embedding
  4. Vectorizer updates row with embedding value

​
Samples

​
Basic usage

Store embeddings in a column for pre-chunked data: 
SELECT ai.create_vectorizer(
    'product_descriptions'::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
);

​
With specific embedding model

SELECT ai.create_vectorizer(
    'short_text_data'::regclass,
    destination => ai.destination_column('text_embedding'),
    loading => ai.loading_column('text'),
    embedding => ai.embedding_ollama('nomic-embed-text', 768),
    chunking => ai.chunking_none()
);

​
Arguments

NameTypeDefaultRequiredDescription
embedding_columnNAME-✔Name of the column to add to the source table for storing embeddings

​
Returns

A JSON configuration object for use in create_vectorizer().

​
Important notes

  • Chunking must be disabled: Use chunking => ai.chunking_none() when using column destination
  • One embedding per row: This approach cannot handle multiple chunks per source row
  • Best for short text: Ideal when text is already chunked or naturally short (< 512 tokens)