Designing Next‑Gen Recommendation and Search with Multi‑Agent AI Architecture

Based on a technical presentation by Alibaba Cloud AI Search lead Xing Shaomin, the article first outlines the challenges of serving high‑concurrency, multimodal data and complex multi‑hop queries. It then describes the evolution of the Agentic Retrieval‑Augmented Generation (RAG) architecture from a single‑agent design to a multi‑agent system that coordinates planning, retrieval, and generation modules to achieve precise intent understanding.

The multi‑route retrieval chain combines vector, text, database, and graph recall strategies. By mixing these sources, coverage and accuracy improve significantly. The author also details GPU‑accelerated indexing and query stages, providing quantitative comparisons that demonstrate substantial speed‑up and reduced latency.

In the Huawei Noah section, the evolution from deep‑learning‑based recommenders to large‑language‑model (LLM)‑driven approaches is examined. Core challenges such as noisy implicit feedback, limited semantic understanding, and difficulty mining user intent are addressed by integrating LLMs as feature enhancers and recommendation components. Using factorization prompting and a multi‑expert knowledge adapter (the KAR project), the solution maps semantic knowledge into the recommendation embedding space, achieving a reported 1.5 % AUC lift and positive online A/B‑test results.

The Baidu GRAB (Generative Ranking for Ads) case study presents a generative ranking model that replaces traditional feature‑engineering pipelines. Leveraging the LLM “Scaling Law” and Transformer architecture, the model jointly encodes user behavior and target ads. A novel Q‑Aware RAB causal attention mechanism introduces query‑aware bias for adaptive modeling of complex interactions and temporal signals. The paper further explains the STS two‑stage training algorithm, heterogeneous token representations, dual‑loss stacking, and KV‑Cache optimization for high‑throughput online inference, along with the quantified business benefits after full deployment.

Additional chapters cover SF Tech’s observability research for OpenAI Swarm‑style multi‑agent systems, a practical guide to building vector search and RAG applications with Elasticsearch, and a forward‑looking discussion on the transition from big‑data to big‑model techniques in search and recommendation.