Hologres:AI models and AI function

AI models

Hologres currently supports the following types of LLMs:

• Managed model: Deploy supported AI models with one click directly in the Hologres console—no API key configuration required. Managed models cover a wide range of scenarios including text generation, multimodal processing, vector embedding, translation, image and video generation, and speech-to-text. After deployment, call these models with AI Functions from SQL. This provides a quick, cost-effective way to add AI capabilities to your data warehouse. For more information, see Managed models.

• AI node models:

  • In addition to managed models, Hologres provides built-in models running on dedicated AI nodes for scenarios such as embedding and inference. You can deploy these models based on your business requirements. These models require you to purchase Hologres AI node (GPU/PPU) resources. If you need dedicated GPU resources and full resource isolation, use built-in models on an AI node. For more information, see Hologres model overview.

  • AI node (GPU/PPU): In addition to CPU resources, Hologres offers GPU/PPU resources, known as AI nodes. When you call an AI node model, the system automatically uses these GPU/PPU resources. This allows you to build AI services quickly without purchasing and managing external GPU resources. AI resources are billed separately, and you can purchase the amount you need. For more information, see AI resource pricing and purchasing.

AI function

Hologres provides a series of built-in functions designed for AI scenarios, such as embed, chunk, and rank. You can use them out-of-the-box like regular functions to call LLMs by using pure SQL. This simplifies development by eliminating the need to build and maintain external AI services. For more information, see AI Function.

Hologres AI evolution

• Hologres V4.0: Added support for managed models—one-click deployment in the Hologres console for cost-effective AI development.

• Hologres V4.0: Introduced major upgrades, including support for Object Table to read unstructured data such as PDF, PPT, and image files. The AI Function feature was enhanced to support unstructured data, which enables end-to-end inference and retrieval when combined with capabilities like vector search and full-text search.

• Hologres V3.2: Introduced AI resources (GPU), integrated several mainstream built-in models, and launched the AI Function feature with operators such as embedding, chunk, and gen. This allows you to call models directly using SQL to quickly build AI applications.

With these updates, Hologres has evolved into a one-stop, multi-modal fusion platform for the AI era. By leveraging AI resources, models, AI Function, Dynamic Table, and other features, you can use standard SQL to manage the entire workflow. This workflow includes data parsing, data processing, model invocation, and real-time retrieval, inference, and analytics. You can develop AI applications just as you would for big data without moving data out of the warehouse. This lets you perform multi-modal data retrieval and analysis within a single platform using standard SQL, unlocking the core value of your data.

SQL-based AI capabilities

Hologres provides various AI Function operators for embedding, ranking, and LLM inference. You can develop end-to-end AI services using standard SQL without calling external AI services, effectively lowering the development barrier.

Simplified AI development

Hologres provides AI resources (GPU/PPU), which eliminates the need to deploy separate GPU/PPU machines. It includes a variety of built-in LLMs that support one-click deployment and GPU resource allocation. You can use the AI Function to call models directly, and the function automatically uses GPU resources. This simplifies the initial setup for AI service development, allowing you to focus on core logic while significantly reducing O&M costs. Since data can be processed by LLMs without being exported, development is simpler and data security is enhanced.

Unified multi-modal search

Hologres provides AI-powered, unified multi-modal search across unstructured, semi-structured, and structured data. You can perform point queries, OLAP analytics, search, and AI inference within a single platform. This simplifies the system architecture by offering a single service endpoint and keeping data in the warehouse.

Multi-modal search and analytics

Hologres allows you to read unstructured data using Object Table. You can then process the data and build indexes with Dynamic Table and the AI Function with operators such as embed and chunk. By combining vector search, full-text search, and dual-channel recall for Top-K results, you can build a one-stop multi-modal search and analytics system with standard SQL. For more information, see Best Practice: Build a multi-modal AI data analytics and search system for finance.

Image search by image or text

Images are a crucial AI input with broad applications, such as vehicle and pedestrian detection in autonomous driving, abnormal behavior tracking in video surveillance, and similar product searches in e-commerce. Hologres allows you to read image metadata by using Object Table and then use Dynamic Table to create embeddings and build vector indexes for the images. At query time, you can use vector recall to implement text-to-image and image-to-image searches. For more information, see Best Practice: Build a high-performance image analytics system for autonomous driving.

Real-time inference

The Hologres AI Function, combined with the platform's high-performance query capabilities, lets you quickly implement features such as real-time label classification, sentiment analysis, and real-time inference without using external models. This significantly reduces development costs.