Baidu researchers propose AI search paradigm: multi-proxy framework for intelligent information retrieval

Cognitive and adaptive search engines are required

With the growth of the demand for situational perception and adaptive information retrieval, modern search systems are developing rapidly. As the number of user queries increases and complexity increases, especially those who require hierarchical reasoning, the system is no longer limited to simple keyword matching or document rankings. Instead, they aim to mimic the cognitive behaviors that humans exhibit when collecting and processing information. The transition to a more complex, collaborative approach marks a fundamental shift in how smart systems respond to users.

Limitations of traditional and rag systems

Despite these advances, the current approach still faces critical limitations. Retrieve enhanced generation (RAG) systems, although available for direct question and answer, usually run in rigid pipelines. They struggle with tasks involving conflicting sources of information, contextual ambiguity, or multi-step reasoning. For example, queries that compare the age of historical figures require understanding, calculating and comparing information from separate documents, i.e. more data is required than simple search and generation. In this case, the lack of adaptive planning and strong reasoning mechanisms often lead to shallow or incomplete answers.

The emergence of multi-proxy architecture in search

Several tools have been introduced to improve search performance, including learning-level systems and advanced retrieval mechanisms utilizing large language models (LLM). These frameworks combine features such as user behavior data, semantic understanding, and heuristic models. However, even advanced rag methods, including React and RQ-rag, follow primarily static logic, which limits its ability to effectively reconfigure plans or recover from execution failures. Their reliance on single-post document retrieval and single-proxy execution further limits their ability to handle complex, context-related tasks.

Introduction to Baidu’s AI search paradigm

Baidu researchers have launched a new approach called the “AI Search Paradigm” that aims to overcome the limitations of static single-grid models. It includes a multi-agent framework with four key agents: master, planner, executor, and writer. Each agent has a specific role during the search process. The master coordinates the workflow according to the complexity of the query. Planners conquer complex tasks. Executor management tools are used and task completion. Finally, the author synthesizes the output into a coherent response. This modular architecture enables flexibility and precise task execution that traditional systems lack.

Use directional acyclic graphs for task planning

The framework introduces directional acyclic graphs (DAGs) to organize complex queries into dependent subtasks. Planners select MCP server-related tools to solve each subtask. The executor then iteratively calls these tools, adjusting the query and fallback strategies when the tool fails or the data is insufficient. This dynamic reallocation ensures continuity and integrity. The author evaluates the results, filters inconsistencies and compiles structured responses. For example, in one query, asking who is older than Han and Julius Caesar, the system retrieves birth dates from different tools, performs age calculations and provides results – all of which are coordinated, multi-agent processes.

Qualitative evaluation and workflow configuration

The performance of the new system was evaluated using multiple case studies and comparative workflows. Unlike traditional rag systems running in single-retrieval mode, the AI ​​search paradigm dynamically complements and reflects on each subtask. The system supports three team configurations based on complexity: author-only, including executors, including wills and planner enhancements. For the Emperor’s Age Comparison Query, the planner breaks down the task into three substeps and assigns the tools accordingly. The final output points out that Emperor Han lived for 69 years and Julius Caesar for 56 years, indicating a difference of 13 years, an output that was accurately synthesized in multiple subtasks. Although this article is more about doomed insight than digital performance metrics, it shows a strong improvement in user satisfaction and robustness across tasks.

Conclusion: Going towards scalable multi-agent search intelligence

In summary, this study proposes a modular, proxy-based framework that enables search systems to go beyond file retrieval and mimic human-style reasoning. The AI ​​search paradigm represents a significant advancement through merger of real-time planning, dynamic execution and coherent synthesis. It not only addresses current limitations, but also provides the foundation for structured collaboration-driven scalable, trustworthy search solutions between smart agents.

Check Paper. All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.

Nikhil is an intern consultant at Marktechpost. He is studying for a comprehensive material degree in integrated materials at the Haragpur Indian Technical College. Nikhil is an AI/ML enthusiast and has been studying applications in fields such as biomaterials and biomedical sciences. He has a strong background in materials science, and he is exploring new advancements and creating opportunities for contribution.