How Agentic Architectures Power Next‑Gen Recommendation and Search Systems

This piece compiles technical deep‑dives from three industry leaders—Alibaba Cloud AI Search, Huawei Noah, and Baidu—into a cohesive analysis of how agentic architectures, large language models (LLMs), and generative ranking are reshaping recommendation and search systems.

Alibaba Cloud AI Search Agentic RAG : The authors describe the challenges of high‑concurrency, multimodal data, and complex multi‑hop queries. They trace the evolution from a single‑agent to a multi‑agent system, where planning, retrieval, and generation modules cooperate to interpret intricate intents. A multi‑path retrieval chain combines vector, text, database, and graph recall to boost coverage and accuracy. GPU‑accelerated indexing and query stages are benchmarked, showing quantifiable speed‑up, and extensions such as NL2SQL and multimodal search are outlined.

Huawei Noah Recommendation Evolution : The article reviews the shift from deep‑learning recommenders to LLM‑augmented and AI‑agent‑driven pipelines. It identifies core pain points—noisy implicit feedback, weak semantic understanding, and intent mining difficulty. Two recommendation paradigms (list‑wise and conversational) are compared, and the integration of LLMs as feature enhancers is detailed. Using factorized prompting and a multi‑expert knowledge adapter, semantic knowledge is efficiently mapped into the recommendation embedding space; the multi‑expert network balances text feature dimensionality with real‑time constraints. Reported results include a 1.5 % AUC lift and supporting online A/B test data.

Baidu GRAB (Generative Ranking for Ads) : To overcome DLRM bottlenecks in feature‑engineering marginal gains, sequence representation loss, and limited generalization, Baidu adopts a generative ranking model inspired by LLM scaling laws and Transformer architecture. User behavior and target ads are jointly encoded in a unified representation space for end‑to‑end sequence generation. The Q‑Aware RAB causal attention mechanism introduces query‑aware bias for adaptive modeling of complex interactions and temporal signals. The paper further details a two‑stage STS training algorithm, heterogeneous token representations, a dual‑loss stacking strategy, and KV‑Cache optimizations that sustain high‑concurrency online inference. Quantitative business impact after full deployment is provided.

The analysis is part of the ebook “Agent Architecture and Practice: Building the Next‑Generation Recommendation and Search Systems,” which also lists additional chapters on multi‑agent interaction for AI‑for‑good, knowledge discovery with LLM agents, observability of OpenAI Swarm, and frontier explorations from big data to large models.