Choose a natural language processing function

This document provides a comparison of the natural language processing functions available in BigQuery ML, which are ML.GENERATE_TEXT , ML.TRANSLATE , and ML.UNDERSTAND_TEXT .

ML.GENERATE_TEXT can do tasks that the other two functions can do as well, as shown in the following image:

You can use the information in this document to help you decide which function to use in cases where the functions have overlapping capabilities.

At a high level, the difference between these functions is as follows:

• ML.GENERATE_TEXT is a good choice for performing customized natural language processing (NLP) tasks at a lower cost. This function offers more language support, faster throughput, and model tuning capability, and also works with multimodal models.
• ML.TRANSLATE is a good choice for performing translation-specific NLP tasks where you need to support a high rate of queries per minute.
• ML.UNDERSTAND_TEXT is a good choice for performing NLP tasks supported by the Cloud Natural Language API .

Supported models

Supported models are as follows:

• ML.GENERATE_TEXT : you can use a subset of the Vertex AI Gemini or PaLM model to generate text. For more information on supported models, see the ML.GENERATE_TEXT syntax .
• ML.TRANSLATE : you use the default model of the Cloud Translation API .
• ML.UNDERSTAND_TEXT : you use the default model of the Cloud Natural Language API .

Supported tasks

Supported tasks are as follows:

• ML.GENERATE_TEXT : you can perform any NLP task. What task the model performs is based on the prompt you specify. For example, to perform a question answering task, you could provide a prompt similar to CONCAT("What are the key concepts in the following article?: ", article_text) . You can also provide context as part of the prompt. For example, CONCAT("context: Only output 'yes' or 'no' to the following question: ", question) .

• ML.TRANSLATE : you can perform the following tasks:

  • TRANSLATE_TEXT
  • DETECT_LANGUAGE

• ML.UNDERSTAND_TEXT : you can perform the following tasks:

  • ANALYZE_ENTITIES
  • ANALYZE_ENTITY_SENTIMENT
  • ANALYZE_SENTIMENT
  • ANALYZE_SYNTAX
  • CLASSIFY_TEXT

Data context

When choosing what function to use, consider whether your data can be analyzed in isolation, or whether it requires additional context to support the analysis. If your data requires additional context, using ML.GENERATE_TEXT with a Vertex AI model is a better choice, as those models allow you to provide context as part of the prompt you submit. Keep in mind that providing additional context as input increases token count and cost.

If your data can be analyzed without other context being considered by the model—for example, translating a string of text without knowing why it was written, then using ML.TRANSLATE or ML.UNDERSTAND_TEXT might be a better choice, provided that the task you want to perform is supported.

Output structure

ML.GENERATE_TEXT consistently returns results in the ml_generate_text_llm_result output column. You can also use your prompt to define your output structure. For example, you can instruct the model to return your result as JSON with custom parent and child fields, and provide an example of how to produce this result.

ML.TRANSLATE and ML.UNDERSTAND_TEXT produce the same output for a given task type for each successful call to the API. Additionally, the output from these functions includes additional metadata about their results. For example, ML.TRANSLATE output includes information about the input language, and ML.UNDERSTAND_TEXT output includes information about the magnitude of the sentiment for sentiment analysis tasks. Generating this metadata is possible with the Vertex AI models, but this requires significant prompt engineering, and isn't likely to provide the same granularity.

Pricing

Pricing is as follows:

• ML.GENERATE_TEXT : For pricing of the Vertex AI models that you use with this function, see Vertex AI pricing . Supervised tuning of supported models is charged at dollars per node hour at Vertex AI custom training pricing .
• ML.TRANSLATE : For pricing of the Cloud AI service that you use with this function, see Cloud Translation API pricing .
• ML.UNDERSTAND_TEXT : For pricing of the Cloud AI service that you use with this function, see Cloud Natural Language API pricing .

Supervised tuning

Supervised tuning support is as follows:

• ML.GENERATE_TEXT : supervised tuning is supported for some models.
• ML.TRANSLATE : supervised tuning isn't supported.
• ML.UNDERSTAND_TEXT : supervised tuning isn't supported.

Multimodality

Multimodality support is as follows:

• ML.GENERATE_TEXT : supports text and text + image input.
• ML.TRANSLATE : supports text input.
• ML.UNDERSTAND_TEXT : supports text input.

Queries per minute (QPM) limit

QPM limits are as follows:

• ML.GENERATE_TEXT : ranges from a 100 to a 1,600 QPM limit in the default us-central1 region , depending on the model used.
• ML.TRANSLATE : 6,000 QPM limit for v3 .
• ML.UNDERSTAND_TEXT : 600 QPM limit .

Token limit

Token limits are as follows:

• ML.GENERATE_TEXT : ranges from a 8,192 to a 24,576 token limit , depending on the model used.
• ML.TRANSLATE : No token limit. However, this function does have a 30,000 bytes limit .
• ML.UNDERSTAND_TEXT : 100,000 token limit .

Supported languages

Supported languages are as follows:

• ML.GENERATE_TEXT : supports the same languages as Gemini or Embeddings , depending on the Vertex AI model you choose for the BigQuery ML remote model endpoint. The PaLM text models support the same languages as the Embeddings models.
• ML.TRANSLATE : supports Cloud Translation API languages .
• ML.UNDERSTAND_TEXT : supports Cloud Natural Language API languages .

Region availability

Region availability is as follows:

• ML.GENERATE_TEXT : available in all Generative AI for Vertex AI regions .
• ML.TRANSLATE : available in the EU and US multi-regions .
• ML.UNDERSTAND_TEXT : available in the EU and US multi-regions.