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

The e‑book "Intelligent Agent Architecture and Practice: Building the Next‑Generation Recommendation and Search System" compiles a series of technical analyses on how AI agents are reshaping recommendation and search pipelines. It covers architecture evolution, multimodal retrieval, large‑language‑model (LLM) integration, and generative ranking techniques.

Alibaba Cloud AI Search – Agentic RAG

Based on a presentation by Alibaba Cloud AI Search lead Xing Shaomin, the chapter details challenges such as high concurrency, multimodal data, and complex multi‑hop queries. It describes the evolution from a single‑agent to a multi‑agent system, where planning, retrieval, and generation modules cooperate to understand and respond to complex intents. The multi‑path retrieval layer mixes vector, text, database, and graph recall strategies to improve coverage and accuracy. GPU‑accelerated indexing and query stages are quantified, showing measurable speed‑up, and extensions like NL2SQL and multimodal search are explained.

Huawei Noah – LLM‑Enhanced Recommendation (KAR Project)

The authors trace the transition from deep‑learning‑based recommenders to LLM‑augmented pipelines, addressing core issues of noisy implicit feedback, weak semantic understanding, and difficulty mining user intent. Two recommendation paradigms—list‑type and conversational—are compared, and the integration of LLMs as feature enhancers is detailed. Using factorization prompting and a multi‑expert knowledge adapter, the KAR project maps semantic knowledge into the recommendation embedding space. The multi‑expert network balances text feature dimensionality with real‑time latency, achieving a reported 1.5% AUC lift and validated by online A/B testing.

Baidu – GRAB Generative Ranking for Ads

GRAB (Generative Ranking for Ads at Baidu) replaces traditional DLRM pipelines that rely on massive discrete features and manual feature engineering. Inspired by LLM scaling laws and the Transformer architecture, it models user behavior and target ads as a unified sequence generation task. A novel Q‑Aware RAB causal attention mechanism introduces query‑aware relative bias to capture complex interactions and temporal signals. The chapter also outlines 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 sustain high‑concurrency online inference. Business impact after full deployment is quantified in the source.

Additional Topics

Other sections explore multi‑agent interaction for AI‑for‑good applications, knowledge discovery with LLM agents, observability research on OpenAI Swarm, and practical guides for building vector search and RAG applications with Elasticsearch. The book provides architecture diagrams, benchmark tables, and detailed implementation notes for each case.