Inside Alibaba Cloud AI Search: Agentic RAG Architecture and Multi‑Agent Techniques

The article introduces an ebook that compiles technical practices from leading AI search and recommendation teams, including Alibaba Cloud, Huawei Noah, and Baidu. Each chapter presents a concrete problem, the engineering challenges faced, and the solutions implemented.

Alibaba Cloud AI Search – Agentic RAG : Based on a talk by Xing Shaomin, the authors describe how the service handles high concurrency, multimodal data, and complex multi‑hop queries. The Agentic RAG architecture progresses from a single‑agent design to a multi‑agent system that separates planning, retrieval, and generation stages. A hybrid retrieval chain mixes vector similarity, textual matching, relational database look‑ups, and graph‑based recall to boost coverage and accuracy. The self‑developed engine is optimized with GPU acceleration for both index construction and query processing; quantization experiments show measurable latency reductions while preserving relevance.

Huawei Noah – Recommendation System Evolution : The chapter reviews the shift from deep‑learning‑based recommenders to large‑language‑model (LLM)‑augmented agents. It details how LLMs are used as feature enhancers and how factorized prompting and a multi‑expert knowledge adapter map semantic knowledge into the recommendation embedding space. The authors report an AUC improvement of 1.5 % and discuss trade‑offs between model size, real‑time latency, and text‑feature dimensionality.

Baidu – GRAB Generative Ranking for Ads : This section explains the design of a generative ranking model that replaces traditional discrete feature pipelines. By adopting the LLM “Scaling Law” and a Transformer backbone, user behavior and ad targets are modeled jointly in an end‑to‑end sequence generation framework. A query‑aware causal attention mechanism (Q‑Aware RAB) captures temporal signals, while a two‑stage STS training schedule and a dual‑loss stacking strategy address training efficiency and over‑fitting. KV‑Cache is employed to sustain high‑throughput online inference.

The ebook’s table of contents lists additional topics such as multi‑agent interaction for AI‑for‑good, large‑model knowledge discovery, observability of OpenAI Swarm‑style systems, and practical guides for building RAG applications with Elasticsearch. Each entry follows the same pattern of problem statement, architectural design, experimental results, and deployment insights.