Configure and call AI models

Text generation functions

• AI_COMPLETE

  • Purpose: A general-purpose text generation function that processes input text according to a given prompt. Suitable for content generation, text rewriting, and general question answering.

  • Usage example: AI_COMPLETE('chat_model', content_string, 'Summarize the input content')

  • Example output: {"result":"Summary of the input content"}

• AI_CLASSIFY

  • Purpose: Classifies input text into specified candidate categories. Separate multiple categories with commas.

  • Usage example: AI_CLASSIFY('chat_model', content_string, 'news,sports,entertainment')

  • Example output: {"category":"sports","confidence":0.95}

• AI_TRANSLATE

  • Purpose: Translates input text from a source language into a target language. When the source language is set to auto, the model detects it automatically.

  • Usage example: AI_TRANSLATE('chat_model', content_string, 'auto', 'en')

  • Example output: {"translated_text":"Hello, world.","detected_language":"zh"}

• AI_SUMMARIZE

  • Purpose: Generates a summary of input text. The third parameter specifies the maximum number of characters allowed in the summary.

  • Usage example: AI_SUMMARIZE('chat_model', content_string, 200)

  • Example output: {"summary":"Summary of the input content"}

• AI_SENTIMENT

  • Purpose: Analyzes the sentiment of input text. The value range of the sentiment score score is -1.0 to 1.0.

  • Usage example: AI_SENTIMENT('chat_model', review_content)

  • Example output: {"score":0.8,"label":"positive","confidence":0.92}

• AI_EXTRACT

  • Purpose: Extracts structured information from unstructured text according to a specified schema.

  • Usage example: AI_EXTRACT('chat_model', content_string, '{"name":"string","age":"integer"}')

  • Example output: {"extracted_json":"{\"name\":\"John\",\"age\":25}"}

• AI_MASK

  • Purpose: Identifies and masks sensitive information of specified types. Separate multiple entity types with commas.

  • Usage example: AI_MASK('chat_model', content_string, 'name,phone,email')

  • Example output: {"masked_text":"Contact: Zhang*, phone: 138****8000","detected_entities":"name,phone"}

Text vectorization functions

• AI_EMBED

  • Purpose: Converts input text into floating-point vectors. Suitable for semantic search, similarity computation, and text clustering.

  • Usage example: AI_EMBED('embedding_model', content_string)

  • Example output: [0.0123, -0.0876, 0.2345, ...]

Multimodal generation functions

• AI_IMAGE_COMPLETE

  • Purpose: Generates text from image binary data and a prompt. Suitable for image captioning, content recognition, and visual question answering.

  • Usage example: AI_IMAGE_COMPLETE('vision_model', image_bytes, 'Describe the product in the image')

  • Example output: "The image shows a pair of white sneakers."

Multimodal vectorization functions

• AI_IMAGE_EMBED

  • Purpose: Converts image binary data into floating-point vectors. Suitable for image-to-image search, image clustering, and multimodal retrieval.

  • Usage example: AI_IMAGE_EMBED('image_embedding_model', image_bytes)

  • Example output: [0.0312, -0.1024, 0.2867, ...]

Create an inference model

VVR 11.7 and later support creating persistent models in the metastore by using CREATE MODEL. When you use built-in models, you do not need to configure endpoint or api-key.

CREATE MODEL qwen_chat
USING openai_compatible
WITH (
  'provider' = 'openai-compat',
  'task' = 'chat/completions',
  'model' = 'qwen3.7-max'
);

CREATE MODEL multimodal_embedding
USING openai_compatible
WITH (
  'provider' = 'dashscope',
  'task' = 'multimodal-embedding',
  'model' = 'qwen3-vl-embedding',
  'content-type' = 'image_url',
  'dimension' = '512'
);

Reuse an existing model

After the model is created, you can reference the Flink SQL model by its full name in a YAML pipeline:

pipeline:
  model:
    - name: chat_model
      reuse-model: vvp.default.qwen_chat

    - name: image_embedding_model
      reuse-model: vvp.default.multimodal_embedding