PolarDB:AI_CreateModel

Syntax

table AI_CreateModel(text model_id, text model_url, text model_provider, text model_type, text model_name, json model_config, regprocedure model_headers_fn, regprocedure model_in_transform_fn, regprocedure model_out_transform_fn);

Parameters

Return values

Returns the AI model creation result in tabular form. The following table describes the result parameters.

Input and output functions

• AI_CreateModel registers the metadata of an AI model in a metadata table but does not call or execute the model. For information about how to call an AI model, see AI_CallModel.

• The input function specified by the model_in_transform_fn parameter converts user-defined input into a JSONB-formatted HTTP request body (required by the AI model). The following example shows the output of the input function for the DeepSeek-R1-Distill-Qwen-7B model deployed on PAI:

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

  The request body must contain the model and prompt parameters. You can define the input function 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;
      $function$;

• The output function specified by the model_out_transform_fn parameter converts the mode' s output, which is usually in the JSON format, into the required format. The following sample code shows the actual result in the JSON format returned by a known AI model:

  {
      "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
      }
  }

  Only the choices[0].text part from the returned result is required. The parameter must be of the JSONB type. You can define the output function as follows:

  CREATE OR REPLACE FUNCTION ai_text_out_fn(model_id text, response_json jsonb)
      RETURNS jsonb
      AS $$
      SELECT (response_json->'choices'->0->'text')::jsonb
      $$ LANGUAGE sql IMMUTABLE;

Examples

The following procedure uses the DeepSeek-R1-Distill-Qwen-7B model deployed on PAI as an example to describe how to register a model.

1. Create an input function for the model.

   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;
       $function$;

2. Create an output function for the model.

   CREATE OR REPLACE FUNCTION ai_text_out_fn(model_id text, response_json jsonb)
       RETURNS jsonb
       AS $$
       SELECT (response_json->'choices'->0->'text')::jsonb
       $$ LANGUAGE sql IMMUTABLE;

3. Register the metadata of the model in a metadata table. Complete the following configurations:

   • Set model_url to the VPC endpoint of the model deployed on PAI with the invocation method /v1/completions appended.

   • Set token in model_config to the token required to access the AI model.

   • Set model_in_transform_fn to the ai_text_in_fn function.

   • Set model_out_transform_fn to the ai_text_out_fn function.

   SELECT polar_ai.AI_CreateModel('my_text_pai_model', '<EAS Endpoint>/v1/completions','Alibaba','EAS LLM','DeepSeek-R1-Distill-Qwen-32B','{"author_type": "token", "token": "<EAS Token>"}', NULL, 'ai_text_in_fn'::regproc, 'ai_text_out_fn'::regproc);

4. View the result. For more information, see Return values.

                     ai_createmodel
   -------------------------------------------------------------
   (1,my_text_pai_model,polar_ai,DeepSeek-R1-Distill-Qwen-7B,t)

5. (Optional) View the created AI model.

   SELECT * from polar_ai._ai_models;

   Sample result:

    model_seq |    model_id    | model_url | model_provider | model_type | model_name | model_config |  model_headers_fn  |     model_in_transform_fn     |     model_out_transform_fn      
   ---
    1 | my_text_embedding_model | https://dashscope.aliyuncs.com/api/v1/services/embeddings/text-embedding/text-embedding | Alibaba | General-purpose text embedding model | 
   text-embedding-v2    | {"token": "20E0BE9E5438xxxxxxxxxxxxxxxxxxxx", "author_type": "token"} |  - | ai_text_embedding_in_fn(text,text) | ai_text_embedding_out_fn(text,jsonb)