# RestGPT: Connecting Large Language Models with Real-World RESTful APIs

**Authors:** Yifan Song, Weimin Xiong, Dawei Zhu, Wenhao Wu, Han Qian, Mingbo Song, Hailiang Huang, Cheng Li, Ke Wang, Rong Yao, Ye Tian, Sujian Li  
**arXiv:** [2306.06624](https://arxiv.org/abs/2306.06624)  
**Date:** 2023-06-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.06624v2/

## Key points

- Tool-augmented large language models (LLMs) have made significant progress in addressing various tasks
- LLMs often struggle with complex instructions in real-world scenarios
- RestGPT is a framework that connects LLMs with RESTful APIs
- RestGPT utilizes a coarse-to-fine online planning mechanism to enhance task decomposition and API selection
- RestGPT includes an API executor tailored for calling RESTful APIs, enabling meticulous parameter formulation and API response parsing
- RestGPT's performance is evaluated using the RestBench benchmark, consisting of two real-world scenarios with human-annotated instructions and gold solution paths
- Experimental results demonstrate impressive results in complex tasks and strong robustness of RestGPT
- The RestGPT framework and RestBench benchmark are publicly available at https://restgpt.github.io/
- Existing models like HuggingGPT have limitations in terms of task coverage compared to RestGPT's integration with RESTful APIs.

## Summary

Tool-augmented large language models (LLMs) have made significant progress in addressing various tasks. However, these models are often limited to specific tools and struggle with complex instructions in real-world scenarios. To overcome these limitations, this paper introduces RestGPT, a framework that connects LLMs with RESTful APIs, which adhere to the widely adopted REST software architectural style for web service development. RestGPT utilizes a coarse-to-fine online planning mechanism to enhance task decomposition and API selection. It also includes an API executor tailored for calling RESTful APIs, enabling meticulous parameter formulation and API response parsing. The performance of RestGPT is evaluated using RestBench, a high-quality benchmark consisting of two real-world scenarios with human-annotated instructions and gold solution paths. Experimental results demonstrate that RestGPT achieves impressive results in complex tasks and exhibits strong robustness, paving the way towards Artificial General Intelligence (AGI). The RestGPT framework and RestBench benchmark are publicly available at https://restgpt.github.io/. Additionally, this paper highlights the limitations of existing models like HuggingGPT which only cover a limited number of tasks compared to the extensive coverage provided by RestGPT's integration with RESTful APIs.