RestGPT is a project that aims to create a large language model autonomous agent to control real-world applications like movie databases and music players. The project connects large language models with RESTful APIs and tackles challenges related to planning, API calling, and response parsing. To evaluate RestGPT’s performance, the team introduces RestBench, which is a high-quality test set consisting of real-world scenarios with human-annotated instructions and solution paths. The code for RestGPT has been released, and the team is currently working on a demo. RestGPT adopts a planning framework and uses RESTful APIs to execute tasks. There are different modules within RestGPT, including a planner, API selector, executor, caller, and parser. Data used for evaluation is collected from TMDB movie database and Spotify music player scenarios. RestBench includes instructions with varying lengths of solution paths. Instructions and solution paths for TMDB and Spotify examples are provided. The code for running RestGPT is available, along with two scripts for running RestGPT on RestBench. The paper has been released, and if this project is helpful, the authors request to be cited according to the provided citation format.
https://github.com/Yifan-Song793/RestGPT