ai_query function

Applies to: check marked yes Databricks SQL check marked yes Databricks Runtime

Preview

This feature is in Public Preview.

Invokes an existing Databricks Model Serving endpoint and parses and returns its response.

See the following guides for how to use ai_query for different use cases:

Requirements

Note

  • In Databricks Runtime 14.2 and above, this function is supported in Databricks notebooks, including notebooks that are run as a task in a Databricks workflow.

  • In Databricks Runtime 14.1 and below, this function is not supported in Databricks notebooks.

Syntax

To query an endpoint that serves an external model or a foundation model:

ai_query(endpoint, request)

To query a custom model serving endpoint with a model schema:

ai_query(endpoint, request)

To query a custom model serving endpoint without a model schema:

ai_query(endpoint, request, returnType, failOnError)

Arguments

  • endpoint: A STRING literal, the name of a Databricks Foundation Model serving endpoint, an external model serving endpoint or a custom model endpoint in the same workspace for invocations. The definer must have CAN QUERY permission on the endpoint.

  • request: An expression, the request used to invoke the endpoint.

    • If the endpoint is an external model serving endpoint or Databricks Foundation Model APIs endpoint, the request must be a STRING.

    • If the endpoint is a custom model serving endpoint, the request can be a single column or a struct expression. The struct field names should match the input feature names expected by the endpoint.

  • returnType: An expression, the expected returnType from the endpoint. This is similar to the schema parameter in from_json function, which accepts both a STRING expression or invocation of schema_of_json function.

    • In Databricks Runtime 14.2 and above, if this expression is not provided, ai_query() automatically infers the return type from the model schema of the custom model serving endpoint.

    • In Databricks Runtime 14.1 and below, this expression is required for querying a custom model serving endpoint.

  • failOnError: (optional) A boolean literal defaulting to true. This flag indicates whether to include error status in the ai_query response.

  • modelParameters (optional): A struct field that contains chat, completion and embedding model parameters for serving foundation models or external models. These model parameters must be constant parameters and not data dependent. When these model parameters are not specified or set to null the default value is used. With the exception of temperature which has a default value of 0.0, the default values for these model parameters are the same as those listed in Foundation model REST API reference.

  • responseFormat (optional): A JSON string field that specifies the response format you want the model to follow. Three string types of response format are supported:

    • text

    • json_object

    • json_schema

Returns

The parsed response from the endpoint.

  • If failOnError => true, the function returns the same result as the existing behavior, which is the parsed response from the endpoint. The data type of the parsed response is inferred from the model type, the model schema endpoint, or the returnType parameter in the ai_query function.

  • If failOnError => false, the function returns the parsed response and the error status string as a STRUCT object.

    • If the inference of the row succeeded, the errorStatus field is null.

    • If the inference of the row failed due to model endpoint errors, the response field is null.

    • If the inference of the row failed due to other errors, the whole query fails.

  • Depending on the JSON string type specified in responseFormat, the following response is returned:

    • For responseFormat => '{“type”, “text”}', the response is a string such as, “Here is the response”.

    • For responseFormat => '{“type”, “json_object”}', the response is a key-value pair JSON string, such as {“key”: “value”}.

    • For responseFormat => '{“type”, “json_schema”...}', the response is a JSON string. See the Examples.

  • If failOnError => false and you have specified responseFormat, the function returns the parsed response and the error status string as a STRUCT object.

Examples

The following example specifies a json_schema response format:

SELECT
  ai_query(
    "llama-3-1-70b",
    <request>,
    responseFormat =>'{
                        "type": "json_schema",
                        "json_schema":
                          {
                           "name": "research_paper_extraction",
                           "schema":
                             {
                              "type": "object",
                              "properties":
                               {
                                "title": { "type": "string" }
                               }
                             },
                            "strict": true
                          }
                      }'
  )

The following is an example output for the specified json_schema response format:

{"title": "<the title of the paper>"}

The following is a batch inference example using the failOnError and modelParameters with max_tokens and temperature. This example also shows how to concatenate the prompt for your model and the inference column using concat(). There are multiple ways to perform concatenation, such as using ||, concat(), or format_string().

CREATE OR REPLACE TABLE ${output_table_name} AS (
  SELECT
      ${input_column_name},
      AI_QUERY(
        "${endpoint}",
        CONCAT("${prompt}", ${input_column_name}),
        failOnError => false,
        modelParameters => named_struct('max_tokens', ${num_output_tokens},'temperature', ${temperature})
      ) as response
    FROM ${input_table_name}
    LIMIT ${input_num_rows}
)

To query an external model serving endpoint:

> SELECT ai_query(
    'my-external-model-openai-chat',
    'Describe Databricks SQL in 30 words.'
  ) AS summary

  "Databricks SQL is a cloud-based platform for data analytics and machine learning, providing a unified workspace for collaborative data exploration, analysis, and visualization using SQL queries."

To query a foundation model supported by Databricks Foundation Model APIs:

> SELECT *,
  ai_query(
    'databricks-meta-llama-3-1-70b-instruct',
    "Can you tell me the name of the US state that serves the provided ZIP code? zip code: " || pickup_zip
    )
  FROM samples.nyctaxi.trips
  LIMIT 10

Optionally, you can also wrap a call to ai_query() in a UDF for function calling as follows:

> CREATE FUNCTION correct_grammar(text STRING)
  RETURNS STRING
  RETURN ai_query(
    'databricks-meta-llama-3-1-70b-instruct',
    CONCAT('Correct this to standard English:\n', text));
> GRANT EXECUTE ON correct_grammar TO ds;
- DS fixes grammar issues in a batch.
> SELECT
    * EXCEPT text,
    correct_grammar(text) AS text
  FROM articles;

To query a custom model serving endpoint:

> SELECT text, ai_query(
    endpoint => 'spam-classification-endpoint',
    request => named_struct(
      'timestamp', timestamp,
      'sender', from_number,
      'text', text),
    returnType => 'BOOLEAN') AS is_spam
  FROM messages
  LIMIT 10


> SELECT ai_query(
    'weekly-forecast',
    request => struct(*),
    returnType => 'FLOAT') AS predicted_revenue
  FROM retail_revenue


> SELECT ai_query(
    'custom-llama-2-7b-chat',
    request => named_struct("messages",
        ARRAY(named_struct("role", "user", "content", "What is ML?"))),
    returnType => 'STRUCT<candidates:ARRAY<STRING>>')

  {"candidates":["ML stands for Machine Learning. It's a subfield of Artificial Intelligence that involves the use of algorithms and statistical models to enable machines to learn from data, make decisions, and improve their performance on a specific task over time."]}

Example query for setting the DLT channel to preview:

> create or replace materialized view
    ai_query_mv
    TBLPROPERTIES('pipelines.channel' = 'PREVIEW') AS
  SELECT
    ai_query("databricks-dbrx-instruct", text) as response
  FROM
    messages
  LIMIT 10;