Reproduce “TURA: Tool-Augmented Unified Retrieval Agent for AI Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhejun Zhao, Yuchen Li, Alley Liu, Yuehu Dong, Xiaolong Wei, Lixue Zheng, Pingsheng Liu, Dongdong Shen and 3 more
The TURA framework integrates traditional Retrieval-Augmented Generation with agentic tool execution, allowing AI search to access both static web content and dynamic, real-time information. Deployed in Baidu Search, the system increased the session success rate by 8.9% and reduced total issues by 16.7% in A/B testing, demonstrating efficient handling of complex, multi-modal queries.
Key insights
- A unified retrieval-and-execution problem formulation over standardized Model Context Protocol (MCP) servers enables seamless integration of static web content and dynamic, real-time data sources in AI search.
- Combining LLM-based multi-intent query decomposition with semantic index augmentation via synthetic query generation drastically improves the accuracy of relevant tool and information source retrieval.
- Directed Acyclic Graph (DAG) based task planning effectively reduces execution latency for complex, multi-hop queries by enabling parallel processing of independent sub-tasks.
- Distilling expert knowledge into a smaller agent using curated trajectories and a "train-with-thought, infer-without-thought" paradigm yields highly accurate, low-latency tool executors suitable for production.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.