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

Usage notes

Supported only in VVR 11.7 or later. Requires activation of Flink AI Service (built-in models).

Register a model

After the master account activates Flink AI Service in the target region, you can create a model in built-in model mode.

CREATE TEMPORARY MODEL my_llm
INPUT (prompt String COMMENT 'Input prompt')
OUTPUT (response String COMMENT 'Model output')
WITH (
  'provider' = 'dashscope',
  'task' = 'chat/completions',
  'model' = 'qwen3.5-flash'
);

The endpoint parameter is not required. The system automatically selects the appropriate service endpoint.
The api-key parameter is not required. The system authenticates using the Flink-managed API Key.
The task parameter is required to declare the model task type.

WITH parameters

General

Parameter	Description	Data type	Required	Default value	Remarks
provider	The model service type.	String	Yes	None	Valid values: `openai-compat`, `dashscope`, and `triton`. `dashscope` (recommended): Supports multimodal tasks such as multimodal-embedding. `openai-compat`: Optional for non-multimodal tasks. `triton`: Used to invoke a Triton inference service. For details, see Invoke a Triton inference service.
task	The model task type.	String	Yes	None	Valid values: chat/completions embeddings multimodal-embedding
model	The specific model to invoke on the service side.	String	Yes	None	Select a model based on the task type. For details, see Built-in model list .
max-context-size	The maximum context size for a single request.	Integer	No	None	If the maximum capacity is exceeded, the action defined in context-overflow-action is triggered.
context-overflow-action	The action to take when a request's context exceeds the maximum capacity.	String	No	`truncated-tail`	Valid values: `truncated-tail`: Automatically truncates tokens when the capacity is exceeded, retaining the most recent `max-context-size` tokens. No logs are recorded. `truncated-tail-log`: Automatically truncates tokens from the tail that exceed the capacity, retaining the most recent `max-context-size` tokens. The truncation is logged. `truncated-head`: Truncates the earliest tokens from the head, retaining the most recent `max-context-size` tokens. `truncated-head-log`: Trims the earliest tokens from the head, keeping the latest `max-context-size` tokens. Logs the truncation. `skipped`: Discards the data record directly. No logs are recorded. `skipped-log`: Discards the data and records a log.
error-handling-strategy	The strategy for handling model request errors.	String	No	retry	Valid values: retry: Resend the request. failover: Throw an exception. ignore: Ignore the exception and skip the data record.
retry-num	The number of retry attempts.	int	No	100	Takes effect only when `error-handling-strategy = retry`.
retry-fallback-strategy	The fallback strategy to use after the maximum number of retries is reached.	String	No	failover	Valid values: Takes effect only when `error-handling-strategy` is set to a value other than `retry`.
retry-backoff-strategy	The retry backoff strategy. Defines how the interval between retries is calculated.	String	No	fixed	Valid values: fixed: Fixed interval. exponential: Exponential interval.
retry-backoff-base-interval	The base time interval for the retry backoff.	Duration	No	1 s	-
content-type	The type of input data. Applies to Chat/Completions and Multimodal-Embedding tasks.	String	No	text	Content type for a single input column. Valid values: `text` (default) or `image_url`. Mutually exclusive with `content-types`.
content-types	The content types of the input columns for multimodal models. Applies to Chat/Completions and Multimodal-Embedding tasks.	String	No	None	A semicolon-separated list. Each input column corresponds to one type (for example, `text;image_url`). Supported types include `text`, `image_url`, and `multi_image_urls`. Mutually exclusive with `content-type`. Note Supported only in VVR 11.8 or later. For details, see General invocation.

chat/completions

Text generation tasks use chat/completions. The following parameters are supported:

Parameter	Description	Data type	Required	Default value	Remarks
system-prompt	The system prompt for the request.	String	No	"You are a helpful assistant."	An empty string is supported.
temperature	Controls the smoothness of the probability distribution for each candidate token.	Float	No	None	Valid range: [0, 2). A value of 0 is not recommended and has no practical meaning. A higher temperature flattens the probability distribution, so more low-probability tokens are selected and the output is more diverse. A lower temperature sharpens the distribution, so high-probability tokens are more likely and the output is more deterministic.
top-p	The probability threshold for nucleus sampling.	Float	No	None	A higher value increases randomness, while a lower value increases determinism.
stop	A stop sequence.	String	No	None	The model stops generating content when the specified string is about to be produced.
max-tokens	The maximum number of tokens that the model can generate.	Integer	No	None	Limited by the model's capabilities.
presence-penalty	Controls token repetition.	Double	No	None	Valid range: -2.0 to 2.0. Positive values penalize tokens that have already appeared in the text, making the model more likely to discuss new topics.
n	The number of outputs to generate for each input.	int	No	None	-
seed	A random number seed for the model's response.	Long	No	None	If specified, the model platform attempts deterministic sampling, so repeated requests with the same seed and parameters return the same result whenever possible.
response-format	The format of the return value.	String	No	text	Valid values: text json_object
extra-body	Extra HTTP body for the request.	String	No	None	Must be a JSON-formatted string. For details, see extra-body description.
user-prompt	The user prompt for the request.	String	No	None	Similar to system-prompt, but submitted from the user's role.

embeddings

Text embedding tasks use embeddings. The following parameters are supported:

Parameter	Description	Data type	Required	Default value	Remarks
dimension	The dimension of the output vectors.	Integer	No	None	The supported dimensions depend on the specific model. Common values are 1024, 768, and 512.

multimodal-embedding

Multimodal embedding tasks use multimodal-embedding to convert text, image, or mixed text-and-image inputs into vectors. The following parameters are supported:

Parameter	Description	Data type	Required	Default value	Remarks
dimension	The dimension of the output vectors.	Integer	No	None	The supported dimensions depend on the specific model. Common values are 1024, 768, and 512.

Example 1: Images only

CREATE TEMPORARY MODEL multimodal_embedding_model
INPUT (`input` STRING)
OUTPUT (`content` ARRAY<FLOAT>)
WITH (
  'provider' = 'dashscope',
  'task' = 'multimodal-embedding',
  'model' = 'qwen3-vl-embedding',
  'dimension' = '512',
  'content-type' = 'image_url'
);

Example 2: Text and image

Text-and-image fusion applies only to models whose fusion capability is enabled by enable_fusion, as described in Multimodal fused vectors. Flink determines whether to enable fusion based on the number of input columns of the model. No additional parameter is required.

CREATE TEMPORARY MODEL fusion_embedding_model
INPUT (text_input STRING, image_input STRING)
OUTPUT (embedding ARRAY<FLOAT>)
WITH (
  'provider' = 'dashscope',
  'task' = 'multimodal-embedding',
  'model' = 'qwen3-vl-embedding',
  'dimension' = '512',
  'content-types' = 'text;image_url'
);

extra-body description

The value of extra-body is a JSON-formatted string used to add extra parameters to the model request body. The available parameters depend on the model service provider. The following are common parameters supported by Alibaba Cloud Model Studio (including but not limited to):

Parameter	Type	Description
`top_k`	Integer	The sampling candidate set size. A larger value increases randomness. For details, see Request body.
`enable_thinking`	Boolean	Whether to enable deep thinking mode (applies to models that support thinking, such as Qwen3). For details, see Deep thinking.
`thinking_budget`	Integer	The maximum token length for the thinking process. Takes effect only when `enable_thinking` is `true` . For details, see Limit thinking length.
`translation_options`	Object	Translation parameters for translation models. For details, see Translation capability.
`enable_search`	Boolean	Whether to use internet search results to assist answer generation. Defaults to `false` . For details, see Web search.
`search_options`	Object	Web search strategy configuration. Takes effect only when `enable_search` is `true` . For details, see Set search scale strategy.

Example

CREATE MODEL my_model
USING openai_compatible
WITH (
  'provider' = 'openai-compat',
  'model' = 'qwen3.5-flash',
  'task' = 'chat/completions',
  'extra-body' = '{"enable_thinking": true, "thinking_budget": 4096}'
);

Querying models

In the Data Query editor, run one of the following commands.

List the names of registered models:

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

Show the statement used to create a model:

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

Show the input and output schema of a 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 |
-- +---------+--------+------+----------+

Modifying models

In the Data Query editor, run the following command.

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; -- Renames the model to m1.

Modify a model parameter:

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

Reset a model parameter to its default value:

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

Deleting models

In the Data Query editor, run the following command.

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

Example

DROP MODEL m;