AI Generate Function

The AI_GENERATE function produces a generated text value for each row returned by the query. For every row, the function sends the provided input text and prompt to the specified AI model and returns the model’s response.

Note:

For information about limitations with using AI Functions, see Known Issues with AI Functions.

Use this function for general-purpose text generation tasks such as:

• Summarization
• Classification
• Rewriting or paraphrasing
• Enrichment or explanation
• Label or tag generation
• General or prompt querying

Syntax

AI_GENERATE(<input>, <prompt>, <model_id>)

Parameters

Parameter	Description
input	String Text representing the row-level data provided to the model. Can be a column, a string literal, or an expression. All non-string values must be explicitly cast to string.
prompt	String Instructions describing what the model should generate. Can be a string literal or an expression.
model_id	String literal Identifier of the model asset to use.

Example

This example classifies each review as positive, negative, or neutral.

SELECT 
          AI_GENERATE(  'Review: ' || REVIEW_TEXT || ';
          Rating: ' || CAST(RATING
          AS VARCHAR(5)),  'Given the review and rating,
          classify the review as positive, negative, or neutral. Return only one word in lowercase
          between the options: positive, negative, or neutral.',  'model'.'id'  ) FROM 
          REVIEWS;

Best Practices

Keep these best practices in mind when you use the AI_GENERATE function:

• Avoid vague prompts. Write clear and explicit instructions describing exactly what the model should produce.
• Specify the desired output format explicitly to prevent extra or unexpected text.
• Provide descriptive labels for each input value to help the model understand what each piece of data represents.
• Ensure the prompt references the same labels used in the input.
• Use punctuation and separators to clearly distinguish between multiple input values.
• Cast non-string expressions / columns to string.
• Use consistent and predictable labels across queries to improve reliability.
• Keep input text concise and focused to improve model quality and reduce unnecessary processing.