ApsaraDB for SelectDB：AI 函數

函數支援

AI 配置相關參數

ApsaraDB for SelectDB 通過資源機制集中管理 AI API 訪問，保障密鑰安全與許可權可控。 現階段可選擇的參數如下：

廠商支援

目前直接支援的廠商有：OpenAI、Anthropic、Gemini、DeepSeek、Local、MoonShot、MiniMax、Zhipu、Qwen、Baichuan。

若有不在上列的廠商，但其 API 格式與 OpenAI/Anthropic/Gemini相同的， 在填入參數ai.provider_type時可直接選擇三者中格式相同的廠商。 廠商選擇只會影響 SelectDB 內部所構建的 API 的格式。

快速上手

1. 配置 AI 資源。

   例 1：

   CREATE RESOURCE 'openai_example'
   PROPERTIES (
       'type' = 'ai',
       'ai.provider_type' = 'openai',
       'ai.endpoint' = 'https://api.openai.com/v1/responses',
       'ai.model_name' = 'gpt-4.1',
       'ai.api_key' = 'xxxxx'
   );

   例 2：

   CREATE RESOURCE 'deepseek_example'
   PROPERTIES (
       'type'='ai',
       'ai.provider_type'='deepseek',
       'ai.endpoint'='https://api.deepseek.com/chat/completions',
       'ai.model_name' = 'deepseek-chat',
       'ai.api_key' = 'xxxxx'
   );

2. （可選）設定預設資源。

   SET default_ai_resource='ai_resource_name';

3. 執行 SQL 查詢。

   case1:

   假設存在如下資料表，表中儲存了與資料庫相關的文檔內容：

   CREATE TABLE doc_pool (
       id  BIGINT,
       c   TEXT
   ) DUPLICATE KEY(id)
   DISTRIBUTED BY HASH(id) BUCKETS 10
   PROPERTIES (
       "replication_num" = "1"
   );

   若需篩選與 SelectDB 相關性最高的 10 條記錄，可採用如下查詢：

   SELECT
       c,
       CAST(AI_GENERATE(CONCAT('Please score the relevance of the following document content to SelectDB, with a floating-point number from 0 to 10, output only the score. Document:', c)) AS DOUBLE) AS score
   FROM doc_pool ORDER BY score DESC LIMIT 10;

   該查詢將利用 AI 產生每條文檔內容與 SelectDB 的相關性評分，並按得分降序篩選前 10 條結果。

   +---------------------------------------------------------------------------------------------------------------+-------+
   | c                                                                                                             | score |
   +---------------------------------------------------------------------------------------------------------------+-------+
   | SelectDB is a lightning-fast MPP analytical database that supports sub-second multidimensional analytics. |   9.5 |
   | In SelectDB, materialized views can automatically route queries, saving significant compute resources.           |   9.2 |
   | SelectDB's vectorized execution engine boosts aggregation query performance by 5–10×.                            |   9.2 |
   | SelectDB Stream Load supports second-level real-time data ingestion.                                      |   9.2 |
   | SelectDB cost-based optimizer (CBO) generates better distributed execution plans.                                |   8.5 |
   | Enabling the SelectDB Pipeline execution engine noticeably improves CPU utilization.                             |   8.5 |
   | SelectDB supports Hive external tables for federated queries without moving data.                                |   8.5 |
   | SelectDB Light Schema Change lets you add or drop columns instantly.                                             |   8.5 |
   | SelectDB AUTO BUCKET automatically scales bucket count with data volume.                                         |   8.5 |
   | Using SelectDB inverted indexes enables second-level log searching.                                              |   8.5 |
   +---------------------------------------------------------------------------------------------------------------+-------+

   case2:

   以下表類比在招聘時的候選人簡曆和職業要求

   CREATE TABLE candidate_profiles (
       candidate_id INT,
       name         VARCHAR(50),
       self_intro   VARCHAR(500)
   )
   DUPLICATE KEY(candidate_id)
   DISTRIBUTED BY HASH(candidate_id) BUCKETS 1
   PROPERTIES (
       "replication_num" = "1"
   );
   
   CREATE TABLE job_requirements (
       job_id   INT,
       title    VARCHAR(100),
       jd_text  VARCHAR(500)
   )
   DUPLICATE KEY(job_id)
   DISTRIBUTED BY HASH(job_id) BUCKETS 1
   PROPERTIES (
       "replication_num" = "1"
   );
   
   INSERT INTO candidate_profiles VALUES
   (1, 'Alice', 'I am a senior backend engineer with 7 years of experience in Java, Spring Cloud and high-concurrency systems.'),
   (2, 'Bob',   'Frontend developer focusing on React, TypeScript and performance optimization for e-commerce sites.'),
   (3, 'Cathy', 'Data scientist specializing in NLP, large language models and recommendation systems.');
   
   INSERT INTO job_requirements VALUES
   (101, 'Backend Engineer', 'Looking for a senior backend engineer with deep Java expertise and experience designing distributed systems.'),
   (102, 'ML Engineer',      'Seeking a data scientist or ML engineer familiar with NLP and large language models.');
   

   可以通過AI_FILTER把職業要求和候選人簡介做語義匹配，篩選出合適的候選人

   SELECT
       c.candidate_id, c.name,
       j.job_id, j.title
   FROM candidate_profiles AS c
   JOIN job_requirements AS j
   WHERE AI_FILTER(CONCAT('Does the following candidate self-introduction match the job description?', 
                   'Job: ', j.jd_text, ' Candidate: ', c.self_intro));

   +--------------+-------+--------+------------------+
   | candidate_id | name  | job_id | title            |
   +--------------+-------+--------+------------------+
   |            3 | Cathy |    102 | ML Engineer      |
   |            1 | Alice |    101 | Backend Engineer |
   +--------------+-------+--------+------------------+

設計原理