How Agentic Architectures Power Next‑Gen Recommendation and Search Systems

Agentic RAG in Alibaba Cloud AI Search

The piece, based on a technical talk by Alibaba Cloud AI Search lead Xing Shaomin, outlines the challenges of high‑concurrency, multimodal data, and complex multi‑hop queries, and presents the Agentic Retrieval‑Augmented Generation (RAG) architecture that evolves from a single‑agent to a multi‑agent system. The design coordinates planning, retrieval, and generation modules to achieve precise intent understanding. It details a multi‑path retrieval chain that mixes vector, text, database, and graph recall to boost coverage and accuracy, and discusses GPU‑accelerated indexing and query quantization, providing comparative performance figures. Extensions such as NL2SQL and multimodal search are also covered, with references to full architecture diagrams and product case studies.

Recommendation System Evolution and LLM Integration (Huawei Noah)

This section reviews the transition of recommendation systems from deep learning to large language models (LLM) and AI agents. It identifies core challenges—noisy implicit feedback, limited semantic understanding, and difficulty mining user intent—and contrasts list‑based and conversational recommendation paradigms. Using Huawei Noah's KAR project as an example, the article explains how LLMs serve as feature enhancers and are integrated via factorized prompting and a multi‑expert knowledge adapter. The adapter’s multi‑expert network balances text feature dimensionality with real‑time constraints. Further analysis covers prompting engineering, fine‑tuning strategies for dialogue recommendation, multi‑capability AI agent coordination, and future cross‑platform ecosystems. Quantitative results show a 1.5% AUC lift and online A/B test data.

GRAB: Baidu Generative Ranking for Ads

The Baidu commercial tech team’s GRAB (Generative Ranking for Ads) model is presented as a solution to the performance bottlenecks of traditional deep learning recommendation models (DLRM). By adopting the LLM “Scaling Law” and Transformer architecture, GRAB models user behavior and target ads in a unified representation space for end‑to‑end generative sequence modeling, eliminating reliance on massive discrete features and manual feature engineering. The article details the Q‑Aware RAB causal attention mechanism that introduces query‑aware relative bias for adaptive modeling of complex interactions and temporal signals. It also describes the STS two‑stage training algorithm for efficiency and over‑fit mitigation, heterogeneous token representations with dual‑loss stacking for model hot‑start, and KV‑Cache optimizations for high‑concurrency online inference. Business impact figures from full deployment are provided.