Baidu Unveils Multi-Agent AI Search Framework for Advanced Information Retrieval

The Need for Smarter Search Engines

Search engines today must handle increasingly complex queries that require reasoning beyond simple keyword matching. Users expect systems to understand context and perform layered information processing similar to human cognition.

Challenges with Traditional Retrieval-Augmented Generation Systems

Current RAG models can answer direct questions but struggle with multi-step reasoning and conflicting data sources. They follow rigid pipelines and lack adaptability, resulting in incomplete or shallow responses for complex queries.

Multi-Agent Architecture: A New Direction

Baidu researchers propose a multi-agent AI Search Paradigm featuring four specialized agents: Master, Planner, Executor, and Writer. This modular setup allows dynamic task division, execution, and synthesis for more accurate and flexible results.

Directed Acyclic Graphs for Task Management

The Planner breaks down complex queries into sub-tasks organized as a Directed Acyclic Graph (DAG). The Executor runs tools iteratively, adapting strategies when needed, while the Writer compiles a coherent final answer. For instance, comparing ages of historical figures involves retrieving data from multiple tools, calculating differences, and delivering a precise response.

Evaluations Highlight Enhanced Robustness

Tests show this system outperforms traditional one-shot retrieval methods by replanning and reflecting on tasks dynamically. Different team configurations enable tailored complexity management, demonstrated by accurately answering queries like age comparisons of Emperor Wu of Han and Julius Caesar.

Advancing Towards Human-Like Search Intelligence

The AI Search Paradigm marks a leap forward by combining real-time planning, adaptive execution, and coherent output generation. This framework paves the way for scalable, trustworthy search systems that emulate human reasoning through collaborative intelligent agents.