BAM-Client

This repository provides an out-of-the-box tool for utilizing OTA's Browser Agent Model (BAM), with enhanced features built on top of existing frameworks. It is intended as a supplementary repo to the model, enabling seamless interaction with web environments through a browser-based agent system.

This repo is forked from:

• Browser-Use (last commit Mar 16 2025) – provides the core browser action framework.
• WebVoyager – contributes the concurrency design and result-saving mechanism.

🧠 About OTA's BAM

The Browser Action Model (BAM) is a lightweight, non-generative model designed by OTA Technology Inc. for intelligent browser-based automation. This repository makes it easy to plug BAM into a fully functional browser action loop with minimal setup.

💻 Demo

📦 Quick Start

Set up conda environment

conda create -n BAM python=3.12
conda activate BAM

Hardware Requirement

Make sure your Dedicated GPU Memory > 20 GB

Please refer to the OTA-v1 for detailed model info.

Run the follwing command to pull OTA-v1 from Hugging Face:

ollama pull hf.co/OTA-AI/OTA-v1

env setup

Setup your virtual environment using pip:

pip install -r requirements.txt

Prepare Tasks

To create your own tasks, follow the format used in the test files under the testcases/ directory. For example, a task in OTA_testdataset_mini.jsonl looks like this:

{"web_name": "Allrecipes", "id": "Allrecipes--4", "ques": "Find a recipe for Baked Salmon that takes less than 30 minutes to prepare and has at least a 4 star rating based on user reviews.", "web": "https://www.allrecipes.com/"}

web_name: the website name you want to visit in this task

id: a unique ID for the task

ques: what you want browser-use to do

web: link to the website

please refer to WebVoyager for more information.

Execute your tasks

Run the following command to start the task:

python run_tasks.py --model-provider ollama --max-concurrent 1 --task_jsonl_path testcases/OTA_testdataset_mini.jsonl

🔧 Improvements Over the Original browser-use Framework

We have extended and improved the browser-use framework with the following key features:

1. Similarity-Based Element Selection
   We integrate similarity search into the web page content symbol space to select only the top-K relevant interactive elements. These are chosen based on their relevance to the agent's next sub-goal, improving both efficiency and model performance.

2. Action History Limiting
   To manage token usage and avoid overwhelming the LLM, we limit the number of previous actions included in the prompt. This helps maintain a concise and effective context for decision-making.

📝 License

This project inherits licensing terms from its upstream forks. Refer to each respective repository for license details.

---

Maintained by OTA Technologies Inc.