AI functions

ai_gen

• Description: Calls an LLM to perform inference on text, images, and videos based on prompts. After deploying the relevant multimodal generation models in the console, you can also use ai_gen for text-to-image, image editing, multi-image blending, text-to-video, and video generation from start frame, start/end frames, or reference images. For model names and parameter details, see Managed models.

  -- Text inference
  SELECT ai_gen([model,] text)
  -- Image inference
  SELECT ai_gen([model,] text, file)
  -- Prompt object input
  SELECT ai_gen([model,] prompt)
  -- Multimodal generation (Example: Specify a model and JSON prompt; 'file' is used for OSS authentication)
  SELECT ai_gen('<model_name>', <prompt_json>::text, to_file('oss://...','oss-xxx-internal.aliyuncs.com','acs:ram::...'));

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • text: (Required) The input prompt. This parameter supports CHAR, VARCHAR, and TEXT types.

  • file: (Required) A FILE type value, such as an image from an Object Table converted to the FILE type. This parameter is supported only in Hologres V4.0 and later.

  • prompt: (Required) A JSON type value returned by the prompt() function.

• Return value

  • Returns the model's response.

  • If the text parameter is NULL, returns NULL.

  • If the text parameter is an empty string (""), returns an empty string ("").

  • If the prompt parameter is NULL, the function reports an error.

• Supported models

  • Hologres AI node built-in models: For text inference, Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B). For image, video, and audio, the qwen-vl-embedding and qwen-vl series are supported.

  • Managed models: For text inference, qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash. For multimodal inference, qwen3-vl-plus. For content generation, qwen-image-2.0-pro, wan2.7-image-pro, the wan2.7 series, and the happyhorse-1.0 series.

• Examples

  • Text inference

    CREATE TABLE questions (
        question TEXT
    );
    
    INSERT INTO questions (question) VALUES
      ('What is artificial intelligence?'),
      ('How can I improve my English speaking skills?'),
      ('What are the key points of a healthy diet?');
    
    SELECT
      question,
      ai_gen('Answer the following question in 20 words: ' || question) AS answer
    FROM
      questions;

    The following result is returned.

                   question                |                                             answer
    -----------------------------------------+-------------------------------------------------------------------------------------------------
     How can I improve my English speaking skills?   | Practice often, mimic pronunciation, build vocabulary, and speak up.
     What are the key points of a healthy diet?      | Eat balanced meals, limit oil, salt, and sugar, eat more fruits and vegetables, and keep regular mealtimes.
     What is artificial intelligence?        | Artificial intelligence is a computer system that simulates human intelligence to learn, reason, perceive, and solve problems.

  • Image inference

    SELECT  ai_gen('jpg_llm','What is in this image?', to_file('oss://****/bd****k/val/images/b9b53753-91a5****.jpg','oss-cn-hangzhou-internal.aliyuncs.com','acs:ram::****' ) )

    The following result is returned.

    ai_gen
    -----
    This image shows a city street scene. A white car with the license plate BTB-9784 is parked on the side of the road. There are several cars on the street, including a yellow taxi. In the background, there are buildings and trees. The weather appears to be overcast, possibly a rainy day. There are also pedestrians and traffic lights on the street.

  • Multimodal generation (images and videos)

    Set prompt to a JSON string that contains business parameters. Pass a to_file function as the third argument to complete OSS authentication. You must pass a valid FILE object to pass the permission check, even when the function does not need to read image or video content. The result is typically a JSON text, with fields such as image_urls and video_url that vary by model. If you configure the output_dir parameter to write results to your OSS bucket, you may incur public network transfer fees. We recommend that you carefully evaluate the potential costs.

    Text-to-image (example)

    SELECT ai_gen(
      'qwen_image_2_pro',
      json_build_object(
        'prompt', 'Example: Generate a product promotion image',
        'parameters', json_build_object('n', 1, 'size', '1024*1024', 'watermark', false)
      )::text,
      to_file('oss://your-bucket/path/placeholder.png', 'oss-cn-hangzhou-internal.aliyuncs.com', 'acs:ram::your-account-id:role/your-role')
    );

    Text-to-video (example)

    SELECT ai_gen(
      'wan26_t2v',
      json_build_object(
        'prompt', 'Example: Show a game character entrance in 10 seconds',
        'parameters', json_build_object(
          'size', '1280*720',
          'prompt_extend', true,
          'duration', 5,
          'shot_type', 'multi'
        )
      )::text,
      to_file('oss://your-bucket/path/placeholder.png', 'oss-cn-hangzhou-internal.aliyuncs.com', 'acs:ram::your-account-id:role/your-role')
    );

    Quick reference for scenarios and models (Actual deployment names are shown in the console. This table is for reference only.)

    Scenario	Example model names
    Text-to-image	qwen-image-2.0-pro, qwen-image-max, qwen-image
    Image editing / multi-image blending	qwen-image-edit, qwen-image-2.0-pro
    Text-to-video	wan2.6-t2v
    Video generation from a start frame, start/end frames, or a reference image	wan2.6-i2v-flash, wan2.2-kf2v-flash, wan2.6-r2v

    For a complete end-to-end pipeline example of game ad video generation, see Best practice: Generate game ad videos with an AI function.

ai_transcribe

• Description: Converts audio files into structured text data. This function is based on Alibaba Cloud Model Studio speech recognition models and supports multiple segmentation granularity modes (none, sentence, speaker, word) for use cases such as meeting transcription, subtitle generation, and speaker diarization. Requires Hologres V4.2 or later.

  For detailed syntax, parameters, and examples, see AI_TRANSCRIBE.

to_file

• Description: Converts a URL to the FILE type.

  select to_file(oss_url, oss_endpoint, oss_rolearn);

• Usage notes

  No model required.

• Parameters

  • oss_url: (Required) TEXT. The path of the OSS file to parse.

  • oss_endpoint: (Required) TEXT. The OSS region endpoint. Only classic network domain names are supported.

  • oss_rolearn: (Required) The RAM role ARN for accessing OSS.

• Return value

  Returns a FILE type value. If the URL path is invalid or the file does not exist, the function reports an error.

• Example

  select to_file('oss://****/bd****k/val/images/b9b53753-91a5****.jpg','oss-cn-hangzhou-internal.aliyuncs.com','acs:ram::****' );

prompt

• Description: Assembles prompts for LLMs, including multimodal prompts.

  SELECT prompt('<template_string>', <expr_1> [ , <expr_2>, ... ])
      FROM <table>;

• Usage notes

  The prompt() function does not support scalar string input. If you have only one string, pass it directly to the LLM function instead of using prompt(). This function is typically used when querying from a table.

• Parameters

  • template_string: (Required) TEXT. The prompt template string. Use placeholders such as {0} and {1} for variables.

  • <expr_1> [ , <expr_2>, ... ]: One or more expression parameters. This parameter supports TEXT, NUMERIC, and FILE types.

• Return value

  In most cases, this function returns a JSON type value in the following format:

  {
    "prompt": "<template_string>",
    "args": ARRAY(<value_1>, <value_2>, ...)
  }

  Special cases:

  • If the template_string parameter is NULL, the function reports an error.

  • If an expression evaluates to NULL, the corresponding placeholder in template_string is replaced with the string 'None'.

  • If all expressions in a row evaluate to NULL, the row is not filtered. Instead, all elements in the args array are populated with 'None'.

• Example

  create table customer_service_konwledge_detail(
    question text, 
    question_summarize text
  );
  
  insert into customer_service_konwledge_detail values
  ('The instance is suddenly experiencing many OOM SQL errors.', 'A backend investigation revealed that an increase in customer traffic exceeded available resources. The customer resolved the issue by scaling up.'),
  ('DataWorks cannot connect to Hologres.', 'This is not our issue. Please contact the DataWorks on-call team.');
  
  -- prompt
  SELECT question, question_summarize, ai_gen(
    prompt('Customer ticket content: {0}, Support response: {1}. Did the response resolve the customer issue? Reply with yes or no only.', question, question_summarize)) 
        from customer_service_konwledge_detail;
       

  The following result is returned.

            question           |                                 question_summarize                                 | ai_gen 
  -----------------------------+------------------------------------------------------------------------------------+--------
   DataWorks cannot connect to Hologres | This is not our issue. Please contact the DataWorks on-call team.                                              | no
   The instance is suddenly experiencing many OOM SQL errors    | A backend investigation revealed that an increase in customer traffic exceeded available resources. The customer resolved the issue by scaling up. | yes
  (2 rows)

ai_parse_document

• Description: Parses unstructured data (PDF, images, Word, PPT, TXT, Markdown) into text.

  SELECT ai_parse_document([model,] input_bytes , input_format [, output_format]);
  SELECT ai_parse_document([model,] file [, input_format, output_format]);

• Parameters

  Parameter	Description
  model	(Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.
  input_bytes	(Required) BYTEA. The binary content of the file to parse.
  file	(Required) FILE. We recommend that you use this function with an Object Table.
  input_format	Optional. The default value is auto. The type is TEXT. Supported file formats include PDF, Word, PPT, TXT, IMAGE, and AUTO. Supported image formats: ["jpg", "jpeg", "png", "tif", "tiff", "bmp"]. If an OSS directory contains documents of multiple types, you can set input_format=auto. The model automatically determines the file type based on the file extension.
  output_format	(Optional) TEXT. The format of the parsed result. Supported formats include JSON and markdown. Default value: JSON.

• Return value

  Returns a TEXT type value. The format is determined by the output_format setting.

  • If output_format=JSON, the output is a TEXT type JSON string. You must explicitly cast it to JSON for use.

  • If parsing fails, the function returns an error description as text instead of reporting an error.

• Supported models: the AI node built-in model ds4sd/docling-models.

• Examples

  • Convert a single PDF file in OSS to text.

    SELECT object_uri, etag, ai_parse_document(to_file ('oss://xxxx-hangzhou/bs_challenge_financial_14b_dataset/pdf', 'oss-cn-hangzhou-internal.aliyuncs.com', 'acs:ram::18xxx:role/xxx'), 'auto', 'markdown') AS doc 
        FROM pdf_bs_challenge_financial_14b_dataset limit 1);

  • To convert unstructured data from an Object Table to text, see Unstructured data (Object Table).

ai_embed

• Description: Computes a fixed-dimensional continuous vector for input text or an image.

  -- Compute text vector
  select ai_embed([model,] content);
  -- Compute image vector
  select ai_embed([model,] file);

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text. This parameter supports CHAR, VARCHAR, and TEXT types.

  • file: (Required) A FILE type value, which is typically the FILE object of an image. This type is supported only in Hologres V4.0 and later.

• Return value

  • If the content parameter is NULL or an empty string, returns NULL.

  • If the file parameter is NULL, returns NULL.

  • The dimension of the returned vector depends on the model used. The supported models and their returned vector dimensions are as follows:

    Model name	Type	Vector dimension	Supported Hologres versions
    clip-ViT-B-32	Image vector	Image patch size: 32 × 32 Parameters: 88 M Dimension: 512	Hologres V4.0 and later
    clip-ViT-B-32-multilingual-v1 Note This model does not support image input. Use clip-ViT-B-32 for images.	Text vector	Image patch size: 32 × 32 Parameters: 88 M Dimension: 512
    clip-ViT-B-16	Image vector	Image patch size: 16 × 16 Parameters: 88 M Dimension: 512
    clip-ViT-L-14	Image vector	Image patch size: 14 × 14 Parameters: 304 M Dimension: 768
    iic/nlp_gte_sentence-embedding_chinese-base	Text vector	512	Hologres V3.2 and later
    iic/nlp_gte_sentence-embedding_chinese-large	Text vector	768
    iic/nlp_gte_sentence-embedding_chinese-small	Text vector	1024
    Qwen/Qwen3-Embedding-0.6B	Text vector	1024
    Qwen/Qwen3-Embedding-4B	Text vector	2560
    Qwen/Qwen3-Embedding-8B	Text vector	4096

• Supported models

  • Hologres AI node built-in models: For text embeddings, the iic/nlp_gte_sentence-embedding_chinese series and the Qwen/Qwen3-Embedding-XB series. For image embeddings, the clip-ViT-B series.

  • Managed models: text-embedding-v4, qwen3-vl-embedding.

• Examples

  • Text embedding

    SELECT ai_embed('Hologres is a one-stop real-time data warehouse engine self-developed by Alibaba, supporting real-time ingestion, updates, processing, and analysis of massive data.');

    The following result is returned.

    ai_embed
    -------
    {-0.020090256, -0.009496426, -0.01584659, ..., -0.057956327}

  • Image embedding

    -- The following example shows how to embed an image from OSS.
    SELECT ai_embed('clip-ViT-B-32', to_file('oss://****', 'oss-cn-hangzhou-internal.aliyuncs.com', 'acs:ram::****'));

Hologres V4.2 and later additionally supports computing image vectors from BYTEA/BLOB binary data. For detailed syntax, managed model list, BYTEA usage examples, and business scenario examples, see AI_EMBED.

ai_rank

• Description: Calculates the relevance score between two texts.

  SELECT ai_rank([model,] source_sentence, sentence_to_compare);

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • source_sentence: (Required) The source text. This parameter supports CHAR, VARCHAR, and TEXT types.

  • sentence_to_compare: (Required) The text to compare with the source_sentence. This parameter supports CHAR, VARCHAR, and TEXT types.

• Return value

  • Returns a FLOAT type relevance score in the range [0, 1]. A higher value indicates higher relevance.

  • If either the source_sentence or sentence_to_compare parameter is NULL, returns 0.

• Supported models: AI node built-in models. Qwen3-series LLMs are recommended (for example, Qwen/Qwen3-32B).

• Example

  SELECT knowledge, ai_rank('What was the revenue of Alibaba in 2024?', knowledge) AS score
    FROM (
        VALUES ('Amazon revenue in 2024 was 638 billion USD'), 
               ('Alibaba revenue in 2024 was 941.168 billion CNY'), 
               ('Alibaba revenue in 2023 was 868.687 billion CNY')
    ) AS knowledge_table(knowledge)
    ORDER BY score DESC;

  The following result is returned.

  knowledge	                  | score
  -----------------------------|-------
  Alibaba revenue in 2024 was 941.168 billion CNY	 |0.899999976
  Alibaba revenue in 2023 was 868.687 billion CNY |0.200000003
  Amazon revenue in 2024 was 638 billion USD	   |0.100000001

ai_chunk

• Description: Splits a long text into chunks.

  SELECT ai_chunk([model,] long_sentence[, chunk_size, chunk_overlap, separators])

• Parameters

  Parameter	Description
  model	(Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.
  long_sentence	(Required) The source text to be chunked. This parameter supports CHAR, VARCHAR, and TEXT types.
  chunk_size	(Optional) INT. The maximum number of characters in each chunk. Default value: 300.
  chunk_overlap	(Optional) INT. Overlap between adjacent chunks to preserve semantic continuity. Default: 50.
  separators	Optional. The separators used to split text into chunks. This parameter is of the TEXT[] type. The default value is ["\n\n", "\n", " ", ""], which is suitable for English text. For Chinese text, we recommend using Chinese separators, such as ["\n\n", "\n", "。", "！", "？", "；", "，", " "].

• Return value

  • Returns a TEXT[] value, which is the list of chunks.

  • If the long_sentence parameter is NULL, returns NULL.

• Supported models: the AI node built-in model recursive-character-text-splitter.

• Example

  SELECT ai_chunk('Hologres is a one-stop real-time data warehouse engine (Real-Time Data Warehouse) self-developed by Alibaba. It supports real-time ingestion, updating, processing, and analysis of massive data, standard SQL (compatible with PostgreSQL protocol and syntax, supporting most PostgreSQL functions), PB-level multidimensional analysis (OLAP) and ad-hoc analysis (Ad Hoc), high-concurrency and low-latency online data services (Serving), and fine-grained isolation for multiple workloads and enterprise-level security capabilities. It is deeply integrated with MaxCompute, Flink, and DataWorks to provide an enterprise-level full-stack data warehouse solution for online and offline integration.',40,10);

  The following result is returned.

  ai_chunk
  ---
  "{\"Hologres is a one-stop real-time data\",\"warehouse engine (Real-Time Data Wareho\",\"use) self-developed by Alibaba. It suppo\",\"orts real-time ingestion, updating, proce\",\"ssing, and analysis of massive data, stand\",\"ard SQL (compatible with PostgreSQL protoc\",\"ol and syntax, supporting most PostgreSQL\",\" functions), PB-level multidimensional an\",\"alysis (OLAP) and ad-hoc analysis (Ad Ho\",\"c), high-concurrency and low-latency onli\",\"ne data services (Serving), and fine-grai\",\"ned isolation for multiple workloads and e\",\"nterprise-level security capabilities. It\",\" is deeply integrated with MaxCompute, Fl\",\"ink, and DataWorks to provide an enterpris\",\"e-level full-stack data warehouse solutio\",\"n for online and offline integration.\"}"

ai_classify

• Description: Classifies input text based on a provided list of category labels.

  SELECT ai_classify([model,] content, labels)

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The text to be classified. This parameter supports CHAR, VARCHAR, and TEXT types.

  • labels: (Required) ARRAY. A list of expected output category labels. The number of labels must be between 2 and 20.

• Return value

  • Returns the matched category label. The return type is TEXT.

  • If the content parameter is NULL, returns NULL.

  • If the content parameter is an empty string (""), returns NULL.

  • If the number of labels is outside the valid range, the function reports an error.

• Supported models

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  CREATE TABLE product_detail(
      product_name TEXT,
      product_desc TEXT
  );
  INSERT INTO product_detail VALUES
  ('iphone','Apple phone'),
  ('p50','Huawei phone'),
  ('x200','vivo phone'),
  ('aaa','Dior dress'),
  ('bbb','Dior pants'),
  ('Sheng Sheng Oolong','Cha Yan Yue Se milk tea'),
  ('sandwich cookie','Oreo cookie');
  
  --Classify text using ai_classify
  SELECT
      product_name,
      ai_classify(product_desc, ARRAY['Electronics', 'Apparel', 'Food & Beverage']) AS category
    FROM
       product_detail
    LIMIT 10;

  The following result is returned.

  product_name	|category
  --------------|------
  aaa	          |Apparel
  iphone	      |Electronics
  Sheng Sheng Oolong	      |Food & Beverage
  p50	          |Electronics
  x200	        |Electronics
  bbb	          |Apparel
  sandwich cookie	      |Food & Beverage

ai_extract

• Description: Extracts specified information from input text and returns the results as a JSON object.

  SELECT ai_extract([model,] content, labels)

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text from which to extract information. This parameter supports CHAR, VARCHAR, and TEXT types.

  • labels: (Required) ARRAY. The labels for the information to extract. The number of labels must be between 1 and 20.

• Return value

  • Returns the extracted information for each label in JSON format.

  • If the content parameter is NULL or an empty string (""), returns NULL.

  • If the number of labels is outside the valid range, the function reports an error.

• Supported models

  • Hologres AI node built-in models: Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B).

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  CREATE TABLE users (
    user_id TEXT,
    resume TEXT
  );
  
  INSERT INTO users (user_id, resume) VALUES
    ('u001', 'Name: Zhang San, Male, 28 years old. Email: zhangsan@example.com, Phone: 1380013****. Extensive work experience.'),
    ('u002', 'Name: Li Si, Female, 35 years old. Phone: 1390013****, Email: lisi@example.com. Has management experience.'),
    ('u003', 'Name: Wang Wu, Male, 25 years old. Email: wangwu@example.com. Phone: 1370013****.');
  
  SELECT
    user_id,
    ai_extract(resume, ARRAY['name','email','phone','gender','age']) AS user_desc_obj
  FROM
    users;

  The following result is returned.

  user_id	|user_desc_obj
  --------|-------------
  u002	  |"{"name":"Li Si","age":"35 years old","gender":"Female","phone":"1390013****","email":"lisi@example.com"}"
  u003	  |"{"name":"Wang Wu","age":"25 years old","gender":"Male","phone":"1370013****","email":"wangwu@example.com"}"
  u001	  |"{"name":"Zhang San","age":"28 years old","gender":"Male","phone":"1380013****","email":"zhangsan@example.com"}"

ai_mask

• Description: Masks specified information in the input text, replacing matches with the [MASKED] placeholder.

  SELECT ai_mask([model,] content, labels)

• Parameters:

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text to be masked. This parameter supports CHAR, VARCHAR, and TEXT types.

  • labels: (Required) ARRAY. The labels for the information to be masked. The number of labels must be between 1 and 20.

• Return value

  • Returns the masked text content.

  • If the content parameter is NULL, returns NULL.

  • If the content parameter is an empty string (""), returns an empty string.

  • If the number of labels is outside the valid range, the function reports an error.

• Supported models

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  SELECT ai_mask(
    'User Wang Xiaoming, ID number: 23030611111111, phone number: 1388888****.',
    ARRAY['ID number', 'phone number']); 

  The following result is returned.

  ai_mask
  -------
  User Wang Xiaoming, ID number: [MASKED], phone number: [MASKED].

ai_fix_grammar

• Description: Corrects grammatical errors in the input text.

  SELECT ai_fix_grammar([model,] content)

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text to correct. This parameter supports CHAR, VARCHAR, and TEXT types.

• Return value

  • Returns the corrected text content.

  • If the content parameter is NULL, returns NULL.

  • If the content parameter is an empty string (""), returns an empty string ("").

• Supported models

  • Hologres AI node built-in models: Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B).

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  SELECT ai_fix_grammar('He dont know what to did.');

  The following result is returned.

  ai_fix_grammar
  --------------
  He doesn't know what to do.

ai_summarize

• Description: Generates a summary of the input text.

  SELECT ai_summarize([model,] content[, max_words])

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text to be summarized. This parameter supports CHAR, VARCHAR, and TEXT types.

  • max_words: (Optional) The target number of words for the summary. The model attempts to generate a result that is close to this value. The default value is 50. If you set this parameter to 0, no limit is applied.

• Return value

  • Returns a summary of the text.

  • If the content parameter is NULL, returns NULL.

  • If the content parameter is an empty string (""), returns an empty string ("").

  • If the value of max_words is less than 0, the function reports an error.

• Supported models

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  SELECT ai_summarize('Hologres is a one-stop real-time data warehouse engine (Real-Time Data Warehouse) self-developed by Alibaba. It supports real-time ingestion, updating, processing, and analysis of massive data, standard SQL (compatible with PostgreSQL protocol and syntax, supporting most PostgreSQL functions), PB-level multidimensional analysis (OLAP) and ad-hoc analysis (Ad Hoc), high-concurrency and low-latency online data services (Serving), and fine-grained isolation for multiple workloads and enterprise-level security capabilities. It is deeply integrated with MaxCompute, Flink, and DataWorks to provide an enterprise-level full-stack data warehouse solution for online and offline integration.', 15);

  The following result is returned.

  ai_summarize
  ------------
  Hologres is Alibaba's self-developed real-time data warehouse engine, supporting real-time processing and multidimensional analysis of massive data.

ai_translate

• Description: Translates input text into a specified language.

  SELECT ai_translate([model,] content, to_lang)

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text to be translated. This parameter supports CHAR, VARCHAR, and TEXT types.

  • to_lang: (Required) The target language code per ISO 639-1.

• Return value

  • Returns the translated text.

  • If the content parameter is NULL, returns NULL.

  • If the content parameter is an empty string (""), returns an empty string ("").

  • If the to_lang parameter value is invalid, the function reports an error.

• Supported models

  • Hologres AI node built-in models: Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B).

  • Managed models: qwen-mt-plus.

• Example

  SELECT ai_translate('Hologres is a one-stop real-time data warehouse engine developed by Alibaba Cloud. It supports real-time data ingestion, updates, processing, and analytics at scale, with standard SQL compatibility (PostgreSQL protocol and syntax). Hologres enables PB-scale multidimensional analysis (OLAP), ad hoc queries, and high-concurrency low-latency online data serving, with fine-grained workload isolation and enterprise-grade security. It is deeply integrated with MaxCompute, Flink, and DataWorks to deliver a unified full-stack data warehouse solution.', 'fr');

  The following result is returned.

  ai_translate
  -----------
  Hologres est un moteur d'entrepot de donnees en temps reel tout-en-un developpe par Alibaba Cloud. Il prend en charge l'ingestion, la mise a jour, le traitement et l'analyse de donnees en temps reel a grande echelle, avec une compatibilite SQL standard (protocole et syntaxe PostgreSQL). Hologres permet l'analyse multidimensionnelle a l'echelle du petaoctet (OLAP), les requetes ad hoc et le service de donnees en ligne a haute concurrence et faible latence, avec une isolation fine des charges de travail et une securite de niveau entreprise. Il est profondement integre a MaxCompute, Flink et DataWorks pour fournir une solution d'entrepot de donnees unifiee et complete.

ai_similarity

• Description: Computes the similarity score between two text inputs.

  SELECT ai_similarity([model,] text1, text2)

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • text1 and text2: (Required) The two texts to be compared for similarity. These parameters support CHAR, VARCHAR, and TEXT types.

• Return value

  • Returns a value of the FLOAT type in the range [0, 1], where a larger value indicates higher similarity. A value of 0 indicates no similarity, and a value of 1 indicates that the two texts are identical.

  • If either the text1 or text2 parameter is NULL, returns 0.

  • If both the text1 and text2 parameters are empty strings (""), returns 1.

  • If one parameter is an empty string ("") and the other is a non-empty string, returns 0.

• Supported models

  • Hologres AI node built-in models: Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B).

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Example

  CREATE TABLE products2 (
      product_name TEXT
  );
  
  INSERT INTO products2 (product_name) VALUES
    ('white shirt'), ('black suit pants'), ('casual top'), ('sports jacket'), ('white dress'),
    ('bluetooth headset'), ('milk chocolate'), ('white top'), ('men''s t-shirt'), ('down jacket');
  
  SELECT product_name FROM products2 
    ORDER BY ai_similarity(product_name, 'white top') DESC LIMIT 5;
  

  The following result is returned.

  product_name
  ----------
  white top
  white shirt
  casual top
  white dress
  men's t-shirt

ai_analyze_sentiment

• Description: Performs sentiment analysis on the input text.

  select ai_analyze_sentiment([model,] content);

• Parameters

  • model: (Optional) The model name used by the AI function. By default, the system automatically assigns the optimal model based on deployed models. To change the model, deploy the target model first, then modify the system table configuration. For details, see Modify the model for an AI function.

  • content: (Required) The input text to be analyzed. This parameter supports CHAR, VARCHAR, and TEXT types.

• Return value

  • Returns the resulting sentiment label as a TEXT value. The returned labels vary by model.

    • Qwen3 series LLMs return one of the following labels: positive, negative, neutral, or mixed. If the input is empty, NULL is returned.

    • The iic/nlp_structbert_sentiment-classification_chinese-base model: Returns the label with the highest probability, which can be positive, negative, or NULL. If the input is empty, it returns NULL.

  • If the content parameter is NULL or an empty string (""), returns NULL.

• Supported models

  • Hologres AI node built-in models: Qwen3-series LLMs are recommended (such as Qwen/Qwen3-32B).

  • Managed models: qwen3.7-max, qwen3.7-plus, qwen3.6-plus, and qwen3.6-flash.

• Examples

  -- Using an LLM:
  SELECT ai_analyze_sentiment('A joyful night of celebration, a time of great success.');
  -- Example output: positive
  
  -- Using the iic/nlp_structbert_sentiment-classification_chinese-base model:
  SELECT ai_analyze_sentiment('A drizzling rain falls on the day of mourning; the traveler on the road is heartbroken.');
  -- Example output: negative

View the mapping between functions and models

Hologres provides the list_ai_function_infos system table, which shows the mapping between AI functions and models. After you deploy a model in the Hologres console, this system table is automatically updated with the model mapping for each AI function. You can call AI functions to invoke the corresponding models.

SELECT * FROM list_ai_function_infos();

The following result is returned.

    function_name     |    model_name    
----------------------+------------------
 ai_embed             | my_gte_embedding
 ai_classify          | my_qwen32b
 ai_extract           | my_qwen32b

Modify the model for an AI function

AI functions have default model mappings. You can modify the model mapped to an AI function using the following methods. After modification, the AI function will invoke the new model.

• Global modification

  SELECT set_ai_function_info('<function_name>', '<model_name>');

  Parameters

  • function_name: The AI function name. You can view AI function names in AI function summary.

  • model_name: The name of a deployed model. Log on to the Hologres console and view deployed model names on the AI Node page.

  Note

  If the specified AI function name or deployed model name does not exist, an error is returned.

  Example

  SELECT set_ai_function_info('ai_embed', 'my_gte_embedding');

• Session-level modification

  After session-level modification, the session-level configuration takes precedence over the global configuration (ai_function_info) when calling AI functions.

  -- Applies only to the current connection
  SET hg_experimental_ai_function_name_to_model_name_mapping='<function_name>:<model_name>[,<function_name1>:<model_name1>]';