Agent Architecture in Action: Building Next‑Gen Recommendation and Search Systems

The piece begins with an in‑depth look at Alibaba Cloud AI Search’s Agentic RAG implementation, which tackles high‑concurrency, multimodal data, and multi‑hop query scenarios. It describes the evolution from a single‑agent to a multi‑agent system, where planning, retrieval, and generation modules cooperate to interpret complex intents. The author explains the multi‑path retrieval chain that mixes vector, text, database, and graph recall to improve coverage and accuracy, and discusses GPU‑accelerated indexing and quantization, as well as extensions such as NL2SQL and multimodal search.

Next, the article examines Huawei Noah’s roadmap from deep‑learning‑based recommenders to large‑language‑model (LLM) and AI‑Agent‑driven solutions. It highlights core challenges like noisy implicit feedback, limited semantic understanding, and difficulty mining user intent. By analyzing list‑style versus conversational recommendation flows, the author details how LLMs serve as feature enhancers within the KAR project, using factorized prompting and multi‑expert knowledge adapters to map semantic knowledge into recommendation embeddings. The design of the multi‑expert network balances feature dimensionality with real‑time latency, and the discussion extends to dialogue‑recommendation prompting, fine‑tuning strategies, and multi‑capability AI‑Agent coordination, reporting an AUC lift of 1.5% and supporting online A/B test data.

The third section focuses on Baidu’s GRAB (Generative Ranking for Ads) model, which replaces traditional DLRM pipelines with an end‑to‑end generative sequence model that jointly encodes user behavior and target ads in a unified representation space. The author breaks down the Q‑Aware RAB causal attention mechanism that introduces query‑aware relative bias for adaptive modeling of complex interactions and temporal signals. Detailed coverage includes the STS two‑stage training algorithm for efficiency and over‑fitting mitigation, heterogeneous token representations, a dual‑loss stacking strategy, and KV‑Cache optimizations that ensure high‑concurrency online inference. Quantitative business benefits observed after full deployment are also presented.

Finally, the article lists the eight chapters of the referenced ebook, covering topics such as multi‑agent interaction for AI benevolence, large‑model knowledge discovery, observability of OpenAI Swarm, Huawei Noah’s recommendation evolution, Baidu GRAB’s generative ranking, Elasticsearch vector search with RAG applications, and broader frontiers from big data to large models in search and recommendation.