Generate Ad Creative with One SQL Using Hologres for Intelligent Creation and Closed‑Loop Analysis

Traditional ad‑creative production suffers from three bottlenecks: low creative capacity that relies on manual effort, a 10‑15 day verification cycle, and a broken data loop between production and placement. Leveraging the rapid growth of large‑model generation, the author presents a solution that integrates non‑structured material collection, tagging, generation, and placement analysis into a unified data system using standard SQL.

Architecture Evolution

Hologres originated as an internal Alibaba real‑time data‑warehouse product (HSAP 1.0) that combined OLAP analysis with online services. With AI capabilities fully integrated—supporting unstructured data management, vector and full‑text search—the product upgraded to HSAP 2.0, positioning itself as a one‑stop AI data‑analysis platform.

The architecture consists of three layers:

Storage layer: Multimodal data are stored in OSS and automatically mapped to an Object Table that records file metadata. Structured and semi‑structured data from MaxCompute, Paimon, and Iceberg are also unified here.

Compute layer: AI Function abstracts large‑model calls as SQL functions such as ai_embed, ai_gen, and ai_translate. These functions invoke Baize models (Embedding, Qwen series, Wan series). Users with dedicated GPU resources can also deploy custom models.

Service layer: Hologres exposes online services for OLAP, vector search, full‑text search, AI inference, and content generation, all accessed through standard SQL.

The core benefits are zero‑code integration (analysts only need SQL), data zero‑migration (both structured tables and OSS files are processed in‑place), real‑time high performance (billion‑scale vector retrieval at sub‑second latency), and flexible low‑cost token‑based billing.

Typical Scenarios

The platform supports three representative use cases: intelligent search & recommendation, real‑time risk control, and intelligent content generation (e.g., automatic marketing copy, video creation, speech synthesis).

Four‑Step Closed‑Loop Pipeline

1. Material Ingestion: Original assets (images, videos, documents) stored in OSS are automatically mapped to the Object Table, extracting metadata and feeding AI Function and a Dynamic Table for downstream processing.

2. Intelligent Tagging: Business teams define a Prompt template (JSON format) that specifies labeling dimensions. By joining Object Table with Prompt and invoking ai_gen, the system incrementally generates structured tags (category, style, scene) without full‑table recomputation.

3. Storyboard & Video Generation: For a 10‑second game‑promotion video, ai_gen calls the Qwen‑Max model to produce a four‑shot storyboard (opening, gameplay, highlight, slogan). The storyboard is then passed to ai_gen with the Wan model, along with reference images, resolution (1280×720), duration, and watermark options, producing a final video stored back in OSS.

4. Automatic Review & Real‑Time Placement Analysis: Generated videos undergo multi‑dimensional compliance, quality, and hotspot evaluation by the Qwen model, which also suggests optimal placement channels (e.g., WeChat, Xiaohongshu, B‑Station). Human reviewers only confirm the final decision. Placement data stream back to Hologres, enabling high‑concurrency writes and immediate analytics using built‑in window functions, funnel analysis, retention, path attribution, and elastic scaling based on traffic load.

Skill → Mem0 Extension

The workflow is packaged as a Hologres Skill. When installed on an Agent (e.g., OpenClaw), users simply provide the material path and generation requirements; the Agent executes the steps: create table → generate image/video → simulate placement → output ROI analysis (e.g., “B‑Station conversion highest, Xiaohongshu needs style adjustment”).

Mem0 adds long‑term memory to agents: it extracts key user information with the Embedding model (1024‑dim vectors), stores it in a Hologres vector store, and retrieves relevant memories in subsequent dialogues, enabling cross‑device consistency for calendar, schedule, and basic profile data.

AI Assistant for Zero‑Code Data‑Warehouse Operations

Hologres AI Assistant bundles expert‑curated Skills covering knowledge Q&A, diagnostics, SQL tuning, data‑warehouse development, and analysis. Users interact via natural language to complete the entire data‑warehouse lifecycle without writing SQL, currently in public beta and free.

Conclusion

The four‑step, SQL‑driven closed loop replaces manual, time‑consuming creative cycles with an automated, data‑driven pipeline that unifies multimodal processing, large‑model inference, and real‑time analytics within a single architecture. The complete workflow and best practices have been open‑sourced as a Hologres Skill for easy replication.