PolarDB:Getting started

Create an extension

Run the following statement as a :

CREATE EXTENSION IF NOT EXISTS polar_ai;

Understand the function for creating custom models

Prepare the following information for the function parameters:

• A deployed and running model service. This example uses the DeepSeek-R1-Distill-Qwen-7B model deployed on Platform for AI (PAI).

  Note

  The deployed model service must be in the same Virtual Private Cloud (VPC) as your PolarDB cluster.

  After deployment, go to Service Details > Call Information to view and record the VPC Call Information. This information includes the Access address and token, which are used to create the model.

• Define model input and output functions.

  There are two invocation methods: Chat and Completions. This topic uses Completions as an example.

  • Model input function:

    When you call the DeepSeek-R1-Distill-Qwen-7B model over HTTP, the request body is as follows:

    {
      "model": "DeepSeek-R1-Distill-Qwen-7B",
      "prompt": "Hello!"
    }

    In the request body, the parameters model and prompt are required as input for the model. The model input function can be defined as follows:

    CREATE OR REPLACE FUNCTION ai_text_in_fn(model_name text, content text)
        RETURNS jsonb
        LANGUAGE plpgsql
        AS $function$
        BEGIN
            RETURN ('{"model": "'|| model_name ||'","prompt":"'|| content ||'"}')::jsonb;
        END;
        $function$;

  • Model output function:

    When you call the DeepSeek-R1-Distill-Qwen-7B model over HTTP, the response body is as follows:

    {
        "id": "8e44xxxx",
        "object": "text_completion",
        "created": 1744355891,
        "model": "DeepSeek-R1-Distill-Qwen-7B",
        "choices": [
            {
                "index": 0,
                "text": " I have a\n\n\n</think>\n\nHello! How can I assist you today? ",
                "logprobs": null,
                "finish_reason": "stop",
                "matched_stop": 151643
            }
        ],
        "usage": {
            "prompt_tokens": 3,
            "total_tokens": 21,
            "completion_tokens": 18,
            "prompt_tokens_details": null
        }
    }

    Because you only need to obtain the content of choices[0].text and the output is a block of text that requires no further processing, you do not need to define a model output function.

Create an AI model

Call the function to create the model. When you create the model, configure the following parameters:

• The model call address model_url is the access address of the model deployed on Platform for AI (PAI), to which you must append the invocation method /v1/completions.

• In the model configuration model_config, the token is the token for the model deployed on Platform for AI (PAI).

• The model input transformation function model_in_transform_fn is the ai_text_in_fn function created in the preceding steps.

SELECT polar_ai.ai_createmodel('my_test_pai_model', '<EAS Endpoint>/v1/completions','Alibaba','EAS large language model','DeepSeek-R1-Distill-Qwen-7B','{"author_type": "token", "token": "<EAS Token>"}', NULL,'ai_text_in_fn'::regproc,NULL);

Create a function to call the AI model

Use the function to create a custom model-calling function:

CREATE OR REPLACE FUNCTION my_text_pai_model_func(model_id text, content text)
    RETURNS text
    AS $$ select (polar_ai.AI_CALLMODEL($1,$2)::jsonb -> 'choices' -> 0 ->> 'text')::text AS result $$
    LANGUAGE 'sql' IMMUTABLE;

Call the model by using SQL

SELECT my_text_pai_model_func('my_test_pai_model', 'Hello');

The following output is returned:

        my_text_pai_model_func        
--------------------------------------
 Hello! How can I assist you today? 
(1 row)