AI_CreateModel

Function description

• AI_CreateModel only registers the model metadata in the metadata table and does not call the model. To call a model, see AI_CallModel.

• The input transform function (model_in_transform_fn) converts input content into the HTTP request content required by the model. The return type must be JSONB. The following example uses the general-purpose text embedding model text-embedding-v2 in Model Studio. Its HTTP request content is:

  {
      "model": "text-embedding-v2",
      "input": {
          "texts": [
          "The wind is high, the sky is vast"
          ]
      },
      "parameters": {
          "text_type": "query"
      }
  }

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

  • model_name: the name of the model to call.

  • content: the text content for which to generate embeddings.

  CREATE OR REPLACE FUNCTION my_text_embedding_in(model_name text, content text)
      RETURNS jsonb
      LANGUAGE plpgsql
      AS $function$
      BEGIN
          RETURN ('{"model": "'|| model_name ||'","input":{"texts":["'|| content ||'"]},"parameters":{"text_type": "query"}}')::jsonb;
      END;
      $function$;

• The output transform function (model_out_transform_fn) converts the model output (typically JSON) into the format required by your application. For example, the model returns the following JSON response:

  {
  "usage": {
      "total_tokens": 7
  },
  "output": {
      "embeddings": [{
            "embedding": [0.004930191827757042, -0.008629344325205105, 0.041976027360927766],
            "text_index": 0
      }]
  },
  "request_id": "317ba0d4-6c08-9c24-8725-eebd445def51"
  }

  Because only the vector content in output.embeddings.embedding is needed and the jsonb format must be retained, the model output function can be defined as:

  • model_id: the parameter type must be TEXT.

  • response_json: the parameter type must be JSONB.

  CREATE OR REPLACE FUNCTION my_text_embedding_out(model_id text, response_json jsonb)
      RETURNS jsonb
      AS $$ select ((((response_json->>'output')::jsonb->>'embeddings')::jsonb)->0->>'embedding')::jsonb as result $$
      LANGUAGE 'sql' IMMUTABLE;

Examples

The following example creates a Text and multimodal embedding model in Model Studio. The registration steps are as follows:

1. Create the input transform function.

   CREATE OR REPLACE FUNCTION my_text_embedding_in(model_name text, content text)
       RETURNS jsonb
       LANGUAGE plpgsql
       AS $function$
       BEGIN
           RETURN ('{"model": "'|| model_name ||'","input":{"texts":["'|| content ||'"]},"parameters":{"text_type": "query"}}')::jsonb;
       END;
       $function$;

2. Create the output transform function.

   CREATE OR REPLACE FUNCTION my_text_embedding_out(model_id text, response_json jsonb)
       RETURNS jsonb
       AS $$ select ((((response_json->>'output')::jsonb->>'embeddings')::jsonb)->0->>'embedding')::jsonb as result $$
       LANGUAGE 'sql' IMMUTABLE;

3. Create the general-purpose text embedding model.

   When creating the model, adjust the following parameters:

   • Model endpoint: modify the model_url domain name by replacing the HTTP endpoint of the general-purpose text embedding model with the PrivateLink endpoint of Model Studio. That is, replace https://dashscope.aliyuncs.com with http://vpc-cn-beijing.dashscope.aliyuncs.com.

   • Model configuration: replace the token in model_config with your API key in Model Studio.

   • Input transform function: set model_in_transform_fn to the ai_text_embedding_in_fn function created in the preceding step.

   • Output transform function: set model_out_transform_fn to the ai_text_embedding_out_fn function created in the preceding step

   SELECT polar_ai.AI_CreateModel('my_text_embedding_model', 'http://vpc-cn-beijing.dashscope.aliyuncs.com/api/v1/services/embeddings/text-embedding/text-embedding','Alibaba','General-purpose text embedding model','text-embedding-v2','{"author_type": "token", "token": "<YOUR_API_KEY>"}', NULL, 'ai_text_embedding_in_fn'::regproc, 'ai_text_embedding_out_fn'::regproc);

   The following result is returned. For more information about the return values, see Return values:

                     ai_createmodel
   --------------------------------------------------------
   (1,my_text_embedding_model,polar_ai,text-embedding-v2,t)

4. (Optional) View the created AI models:

   SELECT * from polar_ai._ai_models;

   Sample output:

    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 | http://vpc-cn-beijing.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)