ML_PREDICT

Applies toBYOCSelf-Managed v3

1 min read

On this page

Limitations
Syntax
Input Parameters
Example
Image Input Example

This topic describes how to use the ML_PREDICT function for AI inference in your Flink jobs.

Limitations

Supported in VERA Engine 4.1 or later.
The throughput of ML_PREDICT operators is subject to the rate limits of your model service provider. When rate limits are reached, the Flink job may experience backpressure at the ML_PREDICT operators, which can lead to timeout errors or job restarts.

Syntax

SQL

1ML_PREDICT(TABLE table_name, MODEL model_name, DESCRIPTOR(input_column_names))

Input Parameters

Parameter	Data type	Description
Parameter	Data type	Description
TABLE table_name	TABLE	The input stream for model inference. This can be a physical table or a view.
MODEL model_name	MODEL	The name of the registered model.
DESCRIPTOR(input_column_names)	N/A	The columns in the input data used for model inference.

Example

The following example registers a sentiment analysis model and uses it to predict sentiment categories for movie reviews.

1. Register the Model

SQL

1CREATE TEMPORARY MODEL ai_analyze_sentiment
2INPUT (input STRING)
3OUTPUT (content STRING)
4WITH (
5    'provider' = 'openai',
6    'endpoint' = '<your-endpoint>',
7    'apiKey' = '<your-key>',
8    'model' = 'gpt-5.1',
9    'system-prompt' = 'Classify the text below into one of the following labels: [positive, negative, neutral, mixed]. Output only the label.'
10);

2. Prepare Test Data

SQL

1CREATE TEMPORARY VIEW movie_comment(id, movie_name, user_comment, actual_label)
2AS VALUES 
3  (1, 'Silent Echo', 'A haunting story that kept me guessing until the end.', 'positive'), 
4  (2, 'The Velvet Gate', 'Nothing special.', 'negative');

3. Run the Prediction

SQL

1SELECT id, movie_name, content as predict_label, actual_label 
2FROM ML_PREDICT(TABLE movie_comment, MODEL ai_analyze_sentiment, DESCRIPTOR(user_comment));

Output Results

The prediction results in the predict_label column match the actual results in the actual_label column.

id	movie_name	predict_label	actual_label
id	movie_name	predict_label	actual_label
1	Silent Echo	POSITIVE	positive
2	The Velvet Gate	NEGATIVE	negative

Image Input Example

Added in VERA Engine 4.6. The following example registers a model for image classification and calls it with an image fetched from S3, passed as a base64-encoded image_url.

1. Register the Model

SQL

1CREATE TEMPORARY MODEL ai_classify_image
2INPUT (input_contents STRING)
3OUTPUT (content STRING)
4WITH (
5    'provider' = 'openai',
6    'endpoint' = '<your-endpoint>',
7    'api-key' = '<your-api-key>',
8    'model' = 'gpt-4.1',
9    'content-type' = 'image_url'
10);

2. Prepare Input Data

Use FETCH_CONTENT to retrieve the image bytes and TO_BASE64 to encode them, computing the column ML_PREDICT will read:

SQL

1CREATE TEMPORARY VIEW product_image(id, input_contents) AS
2SELECT id, TO_BASE64(FETCH_CONTENT(s3_uri))
3FROM (VALUES
4  (1, 's3://my-bucket/products/1.jpg'),
5  (2, 's3://my-bucket/products/2.jpg')
6) AS t(id, s3_uri);

3. Run the Prediction

SQL

1SELECT id, content AS category
2FROM ML_PREDICT(TABLE product_image, MODEL ai_classify_image, DESCRIPTOR(input_contents));

Was this helpful?

Yes No