Model settings - Realtime Compute for Apache Flink - Alibaba Cloud Documentation Center

This topic describes the data definition language (DDL) statements for registering, viewing, modifying, and deleting AI models.

Usage notes

The system supports Alibaba Cloud Model Studio, Platform for AI, and other large model services with OpenAI compatible interfaces.
Large model services deployed on Platform for AI must be in the same region as the real-time computing Flink service.
Only Ververica Runtime (VVR) 11.1 or later is supported.

CREATE MODEL

Register a model

Create a script, and enter a CREATE TABLE command in the SQL editor. Syntax:

CREATE [TEMPORARY] MODEL [catalog_name.][db_name.]model_name
INPUT ( { <physical_column_definition> [, ...n] )
OUTPUT ( { <physical_column_definition> [, ...n] )
WITH (key1=val1, key2=val2, ...)

<physical_column_definition>:
  column_name column_type [COMMENT column_comment]

Clause	Description	Key parameters	Schema requirements	Example
INPUT	Defines the fields, field types, and field order of input data.	`column_name` `column_type` `COMMENT`	Exactly one STRING field is required.	INPUT (`input_text` STRING COMMENT 'User comment')
OUTPUT	Defines the fields, field types, and field order of output data.	`column_name` `column_type` `COMMENT`	Constraints for different task types: chat/completions: Exactly one STRING field is required. embeddings: Exactly one ARRAY<FLOAT> field is required.	OUTPUT (`sentiment_label` STRING COMMENT 'Sentiment label')
WITH	See Parameters.	`provider` `endpoint` `apiKey` `model`	None.	`WITH ('provider'='bailian', 'endpoint'='${ENDPOINT}', 'model'='qwen-turbo', 'apiKey'='${KEY}')`

Examples

Alibaba Cloud Model Studio

CREATE MODEL model_bailian
INPUT (`input` STRING)
OUTPUT (`content` STRING)
WITH (
  'provider'='bailian',
  'endpoint'='<Endpoint>',
  'api-key'='<bailian-key>',
  'model'='qwen3-235b-a22b'
);

Alibaba Cloud Model Studio's endpoint format: <base-url>/compatible-mode/v1/<task>. For example https://dashscope-intl.aliyuncs.com/compatible-mode/v1/chat/completions.

base-url:
Internet access: https://dashscope-intl.aliyuncs.com. To access Alibaba Cloud Model Studio via the internet, enable public access for your Flink workspace. For more information, see Network connectivity selection.
Internal access: Access Alibaba Cloud Model Studio platform through endpoints over internal network. Currently only supports workspaces deployed in the Singapore region.
task: Model task type. Valid values:
- chat/completions
- embeddings

PAI

CREATE MODEL model_pai
INPUT (`input` STRING)
OUTPUT (`embedding` ARRAY<FLOAT>) 
WITH (
  'provider'='bailian',
  'endpoint'='<vpc_endpoint>',
  'api-key'='<Token>',
  'model'='qwen3-235b-a22b'
);

To obtain the VPC endpoint and an API key, see One-click deployment of DeepSeek-V3 and DeepSeek-R1 models.

Log on to the PAI console.
In the left navigation pane, choose Model Gallery > Job Management > Deployment Jobs, and click the corresponding task name.
Click View Invocation Information.
Important
- The VPC invocation address uses http, which needs to be changed to https. The address suffix needs to be appended with /v1/<task>. task: Model task type. The following values are supported:
  - chat/completions
  - embeddings
  Example: https://************.vpc.cn-hangzhou.pai-eas.aliyuncs.com/api/predict/quickstart_deploy_20250722_7b22/v1/chat/completions.
- Token is the value for the api-key parameter.

Parameters

General

Parameter	Description	Data type	Required	Default value	Note
provider	The model service provider.	String	Yes	No default value	Always set it to `bailian`.
endpoint	The API endpoint of Alibaba Cloud Model Studio or other services that offer OpenAI API compatibility for chat, completions, or embeddings.	String	Yes	No default value	For endpoints for Alibaba Cloud Model Studio and PAI, see Examples. For other services, refer to the service's API documentation.
api-key	The API key used to access Alibaba Cloud Model Studio APIs.	String	Yes	No default value	See Obtain API key. In VVR 11.1, the parameter is as `apiKey`.
max-context-size	The maximum context length in a single request.	Integer	No	No default value	This parameter is supported in VVR 11.2 and later. A context window exceeding this limit triggers behavior as defined in `context-overflow-action`.
context-overflow-action	Specifies how to handle a context exceeding the context length limit.	String	No	`truncated-tail`	This parameter is supported in VVR 11.2 and later versions. Valid values: `truncated-tail`: Truncates the context from the end to fit `max-context-size`, without logging the trucation. `truncated-tail-log`: Truncates the context from the end to fit `max-context-size` and logs the truncation. `truncated-head`: Truncates the context from the end to fit `max-context-size`, without logging the truncation. `truncated-head-log`: Truncates the context from the end to fit `max-context-size` and logs the truncation. `skipped`: Aborts the data without logging. `skipped-log`: Aborts the data and logs it.

chat/completions

chat/completions model tasks depend on the following parameters:

Parameter	Description	Data type	Required	Default value	Note
model	The model to call.	String	Yes	No default value	Supports text generation models. Note You are billed based on the LLM selected and the number of input and output tokens.
system-prompt	The system prompt when you request a response.	String	Yes	"You are a helpful assistant."	Parameter name in VVR 11.1: `systemPrompt`
temperature	Controls the randomness of the model's output.	float	No	No default value	Valid value range: [0, 2). Note: The value 0 is meaningless and thus not recommended. A higher temperature makes the less probable token more likely to be selected; a lower temperature increases a model's likelihood of selecting tokens with higher probability.
top-p	Specifies a cumulative probability of tokens considered for sampling.	float	No	No default value	A greater value introduces more randomness to the generated content. Parameter name in VVR 11.1: `topP`
stop	The stop sequence.	String	No	No default value	The model stops generating more text after the specified string appears.
max-tokens	Specifies the maximum length of generated texts.	Integer	No	No default value	Parameter name in VVR 11.1: `maxTokens`

embeddings

Embeddings model tasks depend on the following parameters:

Parameter	Description	Data type	Required	Default value	Note
model	The model to call.	String	Yes	No default value	Supports text embedding models. Note You are billed based on the LLM selected and the number of input and output tokens.
dimension	Controls the output vector dimension.	Integer	No	1024	Valid values:: 1024, 768, 512. Using other values will result in an error.

View models

Create a script and enter the query in the SQL editor.

Show the names of the registered models.

SHOW MODELS [ ( FROM | IN ) [catalog_name.]database_name ];

Show the statement used to register a model.

SHOW CREATE MODEL [catalog_name.][db_name.]model_name;

Show the input and output schema of the model.

DESCRIBE MODEL [catalog_name.][db_name.]model_name;

Example

SHOW MODELS;

-- RESULT
--+------------+
--| model name |
--+------------+
--|          m |
--+------------+

DESCRIBE MODEL m;

-- RESULT
-- +---------+--------+------+----------+
-- |    name |   type | null | is input |
-- +---------+--------+------+----------+
-- | content | STRING | TRUE |     TRUE |
-- |   label | BIGINT | TRUE |    FALSE |
-- +---------+--------+------+----------+

Modify models

Create a script and enter the query in the SQL editor.

ALTER MODEL [IF EXISTS] [catalog_name.][db_name.]model_name {
  RENAME TO new_table_name
  SET (key1=val1, ...)
  RESET (key1, ...)
}

Examples

Rename a registered model.

ALTER MODEL m RENAME TO m1; -- Rename to m1;

Modify model parameters.

ALTER MODEL m SET ('endpoint' = '<Your_Endpoint>'); -- Adjust the endpoint path;

Reset model parameters to default values.

ALTER MODEL m RESET ('endpoint'); -- Reset the endpoint path;

Delete models

Create a script and enter the query in the SQL editor.

DROP [TEMPORARY] MODEL [IF EXISTS] [catalog_name.][db_name.]model_name

Example

DROP MODEL m;