Agent-WebVoyager is an innovative approach to web navigation and data extraction, capable of performing complex browsing tasks without the need for specific APIs and recording progress along the way. Mimicking human-like browsing behavior, WebVoyager navigates the web, interacts with pages, and extracts information, all through visual cues and intelligent decision-making processes.
The project showcases the agent's capability to perform a "meta-webscrape" task, such as browsing X to report Elon Musk's most recent post purely by simulating a few user interactions with the web page. This method stands out by its independence from platform-specific APIs, highlighting a versatile and adaptive web scraping approach. The sky is the limit.
- Human-like Web Navigation: Employs visual cues and page elements for navigation, making the process similar to how a human would browse.
- No API Required: Performs tasks without relying on specific web service APIs, enabling broader applicability across various platforms.
- Intelligent Decision Making: Utilizes a set of defined functions to make decisions, interact with web elements, and navigate through pages. Leverages GPT-4V and LangSmith.
- Visual Task Documentation: Generates a visual path history, documenting each step taken during the task execution.
To set up Agent-WebVoyager, follow these steps:
- Clone the repository:
git clone https://github.com/mrmoxon/Agent-WebVoyager.git - Install the required dependencies:
pip install -r requirements.txt
To run Agent-WebVoyager for a specific task, execute the following command:
python agent_voyage.py --task "Browse Twitter and tell me Musk's most recent tweet."
The agent will perform up to 25 steps to navigate through the web and accomplish the task.
An example task, "Browse Twitter and tell me Musk's most recent tweet.", demonstrates the agent's ability to perform complex web navigation and information extraction without direct API calls. The agent successfully navigates Twitter, finds Elon Musk's profile, and reports the most recent tweet.
The process and steps taken by the agent are documented visually in the path history file: path-history/agent_path_(twitter).png. This file illustrates the agent's navigation path, including interactions and key decisions made along the way.
Using objective = "Could you go to Google Trends and compare 'p(doom)' to 'e/acc'?"
.png)
Note: The example image is a representation. Run the agent to generate current task visual documentation.
Contributions to Agent-WebVoyager are welcome. Please feel free to fork the repository, make changes, and submit pull requests. For major changes, please open an issue first to discuss what you would like to change.