Choose a text generation function

This document provides a comparison of the BigQuery ML ML.GENERATE_TEXT and AI.GENERATE text generation functions. You can use the information in this document to help you decide which function to use in cases where the functions have overlapping capabilities.

Function similarities

The ML.GENERATE_TEXT and AI.GENERATE functions are similar in the following ways:

• Purpose: Generate text by passing a prompt to a large language model (LLM).
• Billing: Incur BigQuery ML charges for data processed. For more information, see BigQuery ML pricing. Incur Vertex AI charges for calls to the LLM. If you are using a Gemini 2.0 or greater model, the call is billed at the batch API rate. For more information, see Cost of building and deploying AI models in Vertex AI.
• Scalability: Process between 1 million and 10 million rows for each a 6-hour query job. Actual throughput depends on factors like the average token length in the input rows. For more information, see Generative AI functions.
• Input data: Support both text and unstructured data from BigQuery standard tables and object tables.

Function differences

Use the following table to evaluate the differences between the ML.GENERATE_TEXT and AI.GENERATE functions:

	ML.GENERATE_TEXT	AI.GENERATE
Function signature	A table-valued function that takes a table as input and returns a table as output.	A scalar function that takes a single value as input and returns a single value as output.
Supported LLMs	Gemini models Partner models such as Anthropic Claude, Llama, and Mistral AI open models	Gemini models
Function output content	Function output content for Gemini models: Generated text Responsible AI (RAI) results Google Search grounding results, if enabled LLM call status Function output content for other types of models: Generated text LLM call status	Generated text Full model response in JSON format LLM call status
Function output format	Generated values are returned in a single JSON column or in separate table columns, depending on the flatten_json_output argument value.	Generated values are returned as fields in a STRUCT object.
User journey	You must create a remote model before using the function.	You can use the function directly, without the need to create a remote model.
Permission setup	You must manually create a BigQuery connection, and grant the Vertex AI User role permission to the service account of the connection. You can skip this step if you are using the BigQuery default connection.	You must manually create a BigQuery connection, and grant the Vertex AI User role permission to the service account of the connection.
Advantages	Allows for more flexible input and output formats.	Easier to integrate into SQL queries.
Extended functions	You can use the AI.GENERATE_TABLE function to generate output that is structured according to a SQL output schema that you specify.	You can use the AI.GENERATE_BOOL, AI.GENERATE_INT, and AI.GENERATE_DOUBLE functions to generate different types of scalar values.