AI Center is a one-stop intelligent engine from EMR Serverless Spark, designed for scenarios that integrate big data and AI. With its two core capabilities—AI Function and model service—you can seamlessly integrate Large Language Model (LLM) capabilities into large-scale data processing flows using only familiar SQL, without writing complex code.
Usage limits
AI Center (Beta) is in public preview. Each Alibaba Cloud account and its RAM sub-accounts share a free usage quota of 1 million tokens. After the total number of consumed tokens exceeds this quota, calls to AI Function will fail.
AI Center will become a paid service on April 8, 2026. For more information, see EMR Serverless Spark AI Center Commercialization Announcement.
Core advantages
SQL-based AI capabilities, zero-code development
Built-in dedicated functions—such as
ai_query, sentiment analysis, and vectorization—eliminate the need to write Python or Java code or manage software development kits (SDKs). Developers use standard SQL to directly call LLMs, embedding complex AI inference seamlessly into existing extract, transform, and load (ETL) flows. This significantly reduces the technical barrier and development costs.Unified service registration, abstracting heterogeneous differences
By default, AI Function directly calls Alibaba Cloud’s latest Qwen3.5-Plus LLM, delivering out-of-the-box, industry-leading inference capability. It also supports flexible business extensions. The model service feature enables one-click registration of models from Alibaba Cloud Model Studio, PAI-EAS, or self-built private models. A unified access layer abstracts underlying heterogeneous differences—including interface protocols and authentication logic—to deliver a “register once, use SQL everywhere” workflow.
In-place data processing, end-to-end intelligent loop
Large-scale vector generation and batch model inference execute in place without moving massive volumes of data. This creates a closed, one-stop loop spanning data cleaning, AI feature engineering, and result write-back. By keeping data within its domain, this approach ensures data security and compliance, eliminates data transfer costs between heterogeneous systems, and greatly simplifies multimodal data processing architecture.
Scenarios
Content understanding
Public opinion and feedback analysis: Automatically identify sentiment—positive or negative—in comments and categorize large volumes of user feedback.
Intelligent ticket routing: Automatically categorize tickets—as complaints, inquiries, or suggestions—based on text content and route them accurately to the appropriate department.
Cross-border business support: Perform end-to-end multilingual translation to quickly generate reports in multiple languages or adapt marketing copy for markets outside China.
General summary generation: Use the Qwen LLM to generate summaries of long documents in batches and extract key information.
Structured data extraction
Key information extraction: Accurately extract predefined fields—such as party A’s name, amount, and date—from contracts, logs, or comments, and output the data directly in JSON format for storage.
Automatic copy polishing: Before data exporting or report generation, automatically correct syntax errors and improper wording to ensure professional external output.
Unstructured to structured data conversion: Convert business rules described in natural language into standardized data records for subsequent SQL analysis.
Semantic retrieval and retrieval-augmented generation (RAG)
Building RAG vector databases: Segment large document sets and convert them into semantic embeddings to provide foundational data support for enterprise knowledge bases.
Intelligent Q&A matching: Calculate semantic similarity between user questions and questions in the knowledge base to achieve high-accuracy auto-replies and retrieval.
Intelligent data deduplication: Identify redundant data with identical meaning but different phrasing—such as duplicate news articles or similar customer service records—to cleanse the dataset.
Data security and compliance
Data masking: Automatically scan text data to identify personal sensitive information—including names, ID card numbers, phone numbers, and bank card numbers—and mask or replace it to ensure data security.
Privacy compliance auditing: Analyze historical data assets in batches to identify unmasked sensitive fields.
Improving SQL development efficiency
Smooth job migration: Automatically convert HiveQL syntax to Spark SQL-compatible syntax, resolving issues related to User-Defined Function (UDF) adaptation and window function standardization.
Query performance tuning: The AI analyzes SQL logic and provides optimization suggestions for predicate pushdown, bucketing strategies, and join operations to improve execution efficiency.
Code commenting and suggestions: Automatically generate comments for complex SQL scripts to help developers quickly understand legacy code logic.