Hologres lets you run large language models (LLMs) directly inside your instance and call them with standard SQL through AI Functions—no separate inference service or extract, transform, and load (ETL) pipeline required. You can use this to build end-to-end AI workflows on your existing data: chunk text, generate embeddings, retrieve relevant context, and run analysis—all within the data warehouse.
Use cases
AI Functions in Hologres support the following scenarios:
Automated content analysis: Classify, summarize, or extract entities from large text tables using SQL queries.
Vector search and RAG pipelines: Generate embeddings and store them in vector fields; combine vector retrieval with full-text search and Online Analytical Processing (OLAP) in a single query.
Multimodal data processing: Read, chunk, vectorize, and analyze multimodal data end-to-end using Dynamic Tables—without moving data out of the warehouse.
BI and data workflow integration: Call models directly from existing SQL-based BI reports and data pipelines. No additional AI engineering expertise is required.
How it works
Deploy a Hologres-managed model from the console with one click, or run a built-in Hologres model on an AI node. Once the model is deployed, call it using AI Functions in SQL. Combine model calls with Dynamic Tables to build automated pipelines that process and analyze data as it arrives.
Model sources
Hologres supports two model sources. The right choice depends on your throughput requirements, cost model, and data residency needs.
Model source | Deployment | Billing | Best for | Operations management | Supported Milvus Version |
Managed model | Deployed in the Hologres console with one click. No API key required. | Pay-as-you-go, aligned with Alibaba Cloud Model Studio pricing. See Managed models billing. | Exploring models, quick integration, and elastic pay-as-you-go usage across text, embedding, image, video, and audio scenarios. | Managed and upgraded automatically by Hologres | Hologres V4.0.40+, V4.1.26+, V4.2.5+, or V5.0+ |
Hologres AI node | Built-in models deployed on dedicated Hologres AI nodes (GPU) | Hologres AI node (GPU) fees only—no separate model invocation fee. See AI resource pricing and purchase. | Core business workloads requiring high queries per second (QPS), low latency, data residency within your domain, and security compliance. | Managed by Hologres AI nodes; administered uniformly within the instance | Hologres V3.2 and later |
How to choose: Start with managed models for fast iteration and pay-as-you-go flexibility. Switch to a Hologres AI node when your workload demands consistent high QPS, strict latency targets, or data must stay within your environment.
Deploy and use models
For deployment steps, see Model deployment.
When the deployment status shows Deployed, connect to your Hologres instance and call the model using AI Functions. For syntax and examples, see AI functions.
What's next
AI functions — Learn how to call deployed models with SQL
High-performance image analytics for autonomous driving — End-to-end example combining AI Functions with real-world multimodal data