Designing Next‑Generation Recommendation and Search Systems with Agentic Architectures

The piece first examines Alibaba Cloud AI Search’s Agentic RAG approach, which tackles high concurrency, multimodal data, and complex multi‑hop queries by evolving from a single‑agent to a multi‑agent system. Planning, retrieval, and generation modules cooperate, and a mixed retrieval pipeline—combining vector, text, database, and graph recalls—boosts query coverage and accuracy. GPU acceleration is quantified for both indexing and query stages.

Next, it reviews Huawei Noah’s recommendation system evolution from deep‑learning models to large language models (LLM) and AI agents. Core challenges such as noisy implicit feedback, limited semantic understanding, and difficulty extracting user intent are addressed by using LLMs as feature enhancers, factorization prompting, and multi‑expert knowledge adapters that balance textual feature dimensionality with real‑time constraints. Reported results include a 1.5% AUC lift and supporting online A/B test data.

The article then details Baidu’s GRAB (Generative Ranking for Ads) model, which replaces traditional DLRM pipelines that suffer from diminishing returns of feature engineering, sequence representation loss, and limited generalization. By applying the LLM “Scaling Law” and Transformer architecture, user behavior and target ads are modeled end‑to‑end in a unified representation space. A Q‑Aware RAB causal attention mechanism captures complex interactions and temporal signals. Additional innovations include a two‑stage STS training algorithm, heterogeneous token representations with dual‑loss stacking, and KV‑Cache for high‑throughput online inference, all contributing to measurable business benefits.

Finally, the e‑book aggregates related chapters on multi‑agent interaction for AI‑for‑good, knowledge discovery with LLM agents, observability of OpenAI Swarm, and Elasticsearch‑based vector search with RAG applications, providing a comprehensive view of cutting‑edge techniques in AI‑driven recommendation and search.