본문으로 건너뛰기
최종 수정: 에

Overview

Description​

AI Function is a built-in function provided by Doris based on Artificial Intelligence (AI) capabilities. Users can directly call AI in SQL queries to perform various intelligent text tasks. AI Function connects to multiple mainstream AI providers (such as OpenAI, Anthropic, DeepSeek, Gemini, Ollama, MoonShot, etc.) through Doris's resource mechanism.

The AI used must be provided externally by Doris and support text analysis.

Configure AI Resource​

Before using AI Function, you need to create a Resource of type AI to centrally manage access information for the AI API.

Example: Create AI Resource​

CREATE RESOURCE "ai_resource_name"
PROPERTIES (
    'type' = 'ai',
    'ai.provider_type' = 'openai',
    'ai.endpoint' = 'https://endpoint_example',
    'ai.model_name' = 'model_example',
    'ai.api_key' = 'sk-xxx',
    'ai.temperature' = '0.7',
    'ai.max_token' = '1024',
    'ai.max_retries' = '3',
    'ai.retry_delay_second' = '1',
    'ai.dimensions' = '1024'
);
Parameter Description​

type: Required, must be ai, used as the type identifier for ai.

ai.provider_type: Required for the general model group. Specifies the external AI provider. Supported providers include OpenAI, Anthropic, Gemini, DeepSeek, Local, MoonShot, MiniMax, Zhipu, QWen, Baichuan, VoyageAI, and Jina. If another provider uses an API compatible with OpenAI, Anthropic, or Gemini, specify the corresponding compatible provider type.

A valid provider type does not imply support for every AI function. For example, VoyageAI and Jina support embedding and cannot serve as text generation models.

ai.endpoint: Required for the general model group. Specifies the AI API endpoint.

ai.model_name: Required for the general model group. Specifies the model name.

ai.api_key: API key for the general model group. Required unless ai.provider_type = local.

ai.temperature: Optional. Controls the randomness of generated content. Accepts a float value between 0 and 1. The default value is -1, which means this parameter is not set.

ai.max_token: Optional. Limits the maximum number of tokens for generated content. The default value is -1, which means this parameter is not set. The default value for Anthropic is 2048.

ai.max_retries: Optional. The maximum number of retries for a single request. The default value is 3.

ai.retry_delay_second: Optional. The delay before retrying, in seconds. The default value on the development branch is 1; in Doris 4.1.5 it is 0.

ai.dimensions: Optional, control the vector dimension of EMBED output. Make sure to confirm that the model filled in ai.model_name supports custom vector dimension before setting, otherwise it may cause model call failure.

Dedicated Embedding Model Configuration (Since 4.1.5)​

Since version 4.1.5, an AI Resource can configure a general model, a text embedding model, and a multimodal embedding model separately. At least one complete group is required. If any connection property or API key is set for a group, that group must include endpoint, provider_type, and model_name, plus api_key for non-LOCAL providers. Missing fields are not inherited from other groups.

Model groupConnection propertiesAPI keyUsage
Generalai.endpoint, ai.provider_type, ai.model_nameai.api_keyText generation and other AI functions; fallback when no dedicated embedding group is configured
Text embeddingai.embed.endpoint, ai.embed.provider_type, ai.embed.model_nameai.embed.api_keyText input to EMBED(); fallback for multimodal input
Multimodal embeddingai.embed.mm.endpoint, ai.embed.mm.provider_type, ai.embed.mm.model_nameai.embed.mm.api_keyJSON multimodal input to EMBED()

Text EMBED() selects the text embedding group first, then the general group. Multimodal EMBED() selects the multimodal embedding group, then the text embedding group, then the general group. The selected provider and model must support the actual media type. A resource with only embedding groups can serve the corresponding EMBED() calls, but cannot serve text generation functions that require a general model.

For example, a resource used only for text embedding does not need a general generation model:

CREATE RESOURCE "text_embed_resource"
PROPERTIES (
    "type" = "ai",
    "ai.embed.provider_type" = "openai",
    "ai.embed.endpoint" = "https://<embedding-service>/v1/embeddings",
    "ai.embed.model_name" = "<embedding-model>",
    "ai.embed.api_key" = "<api-key>"
);

Replace the connection values with your service configuration. Successful creation returns Query OK. Shared settings such as ai.dimensions and retry parameters still come from the resource configuration.

Reasoning Effort: ai.effort (Since 4.1.5)​

Since version 4.1.5, the optional string property ai.effort specifies reasoning effort. By default, it is unset and is not sent to the provider. Use lowercase values, subject to provider validation:

ProviderAccepted values
ANTHROPIClow, medium, high, xhigh, max
GEMINIminimal, low, medium, high
Property validation for other providersnone, minimal, low, medium, high, xhigh, max

Passing property validation does not mean every model supports that effort level. Choose a value supported by your service and model. ai.effort does not replace ai.max_token.

Create, Alter, and Drop Permissions (Since 4.1.5)​

Creating an AI Resource requires global ADMIN privileges. Since version 4.1.5, the FE configuration ai_resource_allowed_user can further restrict the creator's identity. Its default value, *, adds no identity restriction.

An AI Resource newly created by root since version 4.1.5 can only be altered or dropped by root; other accounts are rejected even if they have ADMIN privileges. Resources that existed before the upgrade and lack this marker are not marked retroactively. Other resources retain the existing privilege checks.

Resource Selection and Session Variables​

When users call AI-related functions, resources can be specified in the following two ways:

  • Explicitly specify the resource: directly pass the resource name when calling the function.
  • Implicitly specify the resource: set the Session variable in advance, and the function will automatically use the corresponding resource.

Set Session variable format:

SET default_ai_resource='resource_name';

Function call format:

SELECT AI_FUNCTION([<resource_name>], <args...>);

Resource Selection Priority​

When calling an AI_Function, it determines which resource to use in the following order:

  1. The resource explicitly specified by the user in the call
  2. The global default resource (default_ai_resource)

Example:

SET default_ai_resource='global_default_resource';
SELECT AI_SENTIMENT('this is a test'); -- Uses resource named 'global_default_resource'
SELECT AI_SENTIMENT('invoke_resource', 'this is a test'); --Uses resource named 'invoke_resource'

AI Functions​

Currently supported AI Functions in Doris include:

  • AI_CLASSIFY: Information classification

  • AI_EXTRACT: Information extraction

  • AI_FILTER：Filter information

  • AI_FIXGRAMMAR: Grammar correction

  • AI_GENERATE: Text generation

  • AI_MASK: Masking sensitive information

  • AI_SENTIMENT: Sentiment analysis

  • AI_SIMILARITY: Text semantic similarity comparison

  • AI_SUMMARIZE: Text summarization

  • AI_TRANSLATE: Translation

  • AI_AGG: Perform cross-line aggregation analysis on multiple texts

Examples​

  1. AI_TRANSLATE
SELECT AI_TRANSLATE('resource_name', 'this is a test', 'Chinese');
-- 这是一个测试
  1. AI_SENTIMENT
SET default_ai_resource = 'resource_name';
SELECT AI_SENTIMENT('Apache Doris is a great DBMS.');

For detailed function and usage, please refer to the documentation of each specific function.

Report issue