AutoFlow: Automated Workflow Generation for Large Language Model Agents

Recent advancements in Large Language Models (LLMs) have shown significant progress in understanding complex natural language. However, LLMs still face challenges in generating and executing programming codes accurately. While some efforts have been made to leverage LLMs for code generation, many generated codes are still unable to be executed effectively. In contrast, natural language programs can have a higher executable rate due to their minimal syntax requirements compared to traditional programming languages. As evidence, recent work proposes the CoRE language, utilizing LLMs as interpreters for workflow programming via natural language, achieving a higher valid plan rate compared to baseline methods. However, CoRE requires manual design and may lead to suboptimal solutions. To address these issues, we propose AutoFlow, a framework designed to automatically generate workflows in the CoRE language for solving complex tasks. AutoFlow offers two workflow generation methods: fine-tuning-based and in-context-based methods, making it applicable to both open-source and closed-source LLMs. Our framework produces more robust and reliable workflows than existing code generation methods. Moreover, natural language programming offers greater readability and lower barriers for coders than traditional programming languages. We believe that the automatic generation and interpretation of workflows in natural language represent a promising paradigm for solving complex tasks, particularly with the rapid development of LLMs.

This package is mainly contributed by Zelong Li ([email protected]), Shuyuan Xu ([email protected]), and Yongfeng Zhang ([email protected]). We welcome any issues and requests for model implementation and bug fix.

Citation

@article{li2024autoflow,
  title={AutoFlow: Automated Workflow Generation for Large Language Model Agents},
  author={Li, Zelong and Xu, Shuyuan and Mei, Kai and Hua, Wenyue and Rama, Balaji and Raheja, Om and Wang, Hao and Zhu, He and Zhang, Yongfeng},
  year={2024}
}

Requirements

• Python==3.9
• PyTorch==2.2.2
• transformers==4.40.2
• langchain==0.1.4
• peft==0.7.1

Preparation

0. Clone this repo.

1. Create a conda virtual environment and install the Pytorch matching your CUDA version. For example, for CUDA version 12.1:

conda create -n your_env_name python=3.9
conda activate your_env_name

conda install pytorch==2.2.2 torchvision==0.17.2 torchaudio==2.2.2 pytorch-cuda=12.1 -c pytorch -c nvidia

2. Install necessary packages:

pip install -r requirements.txt
pip install -r requirements_travel.txt

3. Download the OpenAGI data from this Google Drive link, unzip it to the AutoFlow directory and rename it as openagi_data.

4. Download the database and unzip it to the AutoFlow directory (i.e., your/path/AutoFlow) and rename it as travel_database.

5. Make sure you are in the AutoFlow/src folder before running the codes. Otherwise,

cd src

Running Command Examples

(Notice that --model_name can be different from --auto_model_name)

OpenAGI task when using gpt-4-1106-preview as the workflow interpreter LLM:

python auto_main.py 
--flow_name=OpenAGI_Flow.txt
--task=OpenAGI 
--model_name="gpt-4-1106-preview"
--auto_model_name="gpt-4-1106-preview"
--log_file_name=../log/autoagi_gpt4gpt.txt
--output_dir=./gpt4gpt
--auto_flow_name="autoagi_gpt4gpt_Flow.txt"
--auto_epoch=30 
--openai_key="YOUR OPENAI KEY"
--max_round=20

OpenAGI task when using TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ as the workflow interpreter LLM:

python auto_main.py 
--flow_name=OpenAGI_Flow.txt
--task=OpenAGI 
--model_name="TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ"
--auto_model_name="mistralai/Mixtral-8x7B-Instruct-v0.1"
--log_file_name=../log/autoagi_mixtral4mixtral.txt
--output_dir=./mixtral4mixtral
--auto_flow_name="autoagi_mixtral4mixtral_Flow.txt"
--auto_epoch=30 
--openai_key="YOUR OPENAI KEY"
--max_round=20

TravelPlanner task when using gpt-4-1106-preview as the workflow interpreter LLM:

python auto_main.py 
--flow_name=TravelPlanner_Flow_manual.txt 
--tool_name=tools.txt 
--task=TravelPlanner 
--model_name="gpt-4-1106-preview"
--auto_model_name="gpt-4-1106-preview"
--log_file_name=../log/auto_travel_gpt4gpt.txt 
--auto_flow_name=TravelPlanner_Flow_gpt4gpt.txt
--auto_epoch=30 
--openai_key="YOUR OPENAI KEY"
--max_round=100

TravelPlanner task when using TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ as the workflow interpreter LLM:

python auto_main.py 
--flow_name=TravelPlanner_Flow_manual.txt 
--tool_name=tools.txt 
--task=TravelPlanner 
--auto_model_name=gpt-4-1106-preview 
--model_name=TheBloke/Mixtral-8x7B-Instruct-v0.1-GPTQ 
--log_file_name=../log/auto_travel_gpt4mixtral.txt 
--auto_flow_name=TravelPlanner_Flow_gpt4mixtral.txt 
--auto_epoch=30 
--max_round=100

Known Issues

PPOTrainer class in the latest version of trl package (==0.9.4) has a known issue when using multiple GPUs for reinforcement learning.

Generated Workflow

In the AutoFlow/src/info/OpenAGI folder, there are manually designed workflow and automatically generated workflows. The manual workflow is OpenAGI_Flow_manual.txt, and the file name of automatically generated workflows is in the form of OpenAGI_Flow_manual_*4*.txt. For example, OpenAGI_Flow_manual_gpt4mixtral.txt means the workflow is generated by GPT and used for Mixtral as the interpreter LLM.

Reference

• We leveraged the dataset of OpenAGI and TravelPlanner projects and based on CoRE language to implement our experiment.