SafeChat: A Framework for Building Trustworthy Collaborative Assistants

SafeChat is an architecture to create collaborative assistants, also called chatbots and digital assistants, which can serve as safe decision supports to provide reliable, authenticated, information in trust-sensitive domains. It uses a combination of neural (learning-based, including generative AI) and symbolic (rule-based) methods, together called a neuro-symbolic approach, to provide known information in easy-to-use consume forms that are adapted from user interactions (provenance). The chatbots generated are scalable, quick to build and have in-built support for evaluation of trust issues like fairness, robustness and appropriateness of responses. SafeChat is implemented over the open-source, Rasa platform, to create a tool (and hence, an executable framework) but the approach is platform-agnostic.

🎯 Key Features

• For safety

  • Safe design where only responses that are grounded and traceable to an allowed source will be answered (provenance).
  • Supports do-not-respond strategy that can deflect certain user questions which may be harmful if answered.

• For usability

  • Supports automated trust ratings to communicate the chatbot’s expected behavior on dimensions like abusive language and bias.
  • Supports automatic, extractive summarization of long answers that can be traced back to source

• For fast, scalable, development

  • Provides a CSV-driven chatbot building workflow that does not require deep AI expertise, making it accessible to developers with varying levels of AI knowledge and experience.
  • Domain agnostic, scalable architecture. The backend can be extended with CSV-driven web integration.
  • Support for testing, including control and treatment group formation and analysis of results, for randomized control trial (RCT)

📋 Prerequisites

Before you begin, ensure you have the following installed:

• Python
• Anaconda

🚀 Getting Started

1. Clone the repository

   git clone https://github.com/ai4society/trustworthy-chatbot
   cd trustworthy-chatbot

2. Create and activate a conda virtual environment

   conda create --name safechat_venv python==3.8
   conda activate safechat_venv

3. Install required packages

   pip install -r requirements.txt

4. Provide question answer pairs for your use case

   • Create a new data/input/<file-name>.csv file with your desired question answer pairs or use the existing one. Your CSV should have a column called Question and another one called Answer. Optionally, if you would like for your chatbot to give traceable responses to your user, include columns Timestamp with UNIX timestamps corresponding to the date of the response and Source with strings which may be URLs or names of the organization which procured the response.
   • Create a new data/input/DNA.csv that is a single column list of questions that you would like your chatbot to avoid answering. Ensure that the heading of this column is Questions

5. Process your data and configure the chatbot using the RASA Open Source framework

   • Extract the intent for user queries: python code/extract_intent.py By default, this will read data/input/Chat.csv, however, you can specify a different csv file in the data/input directory with the -f or --file argument:python code/extract_intent.py -f Election_QA.csv
     • Saves to data/input/Chat_intent.csv
   • Paraphrase all user queries: python code/paraphraser.py
     • Saves to data/input/paraphrased.json
   • Configure directory for RASA Open Source: python code/configure_rasa.py
     • Saves to Chatbot directory. Deletes existing Chatbot by default, so save in different location if you would like to save the configuration.

6. Train and converse with your chatbot

   • Navigate to chatbot directory: cd Chatbot
   • Train chatbot: rasa train
   • Converse with trained chatbot: rasa shell
   • Optionally, if you would like your chatbot to handle some business-logic for some queries, you may utilize the RASA Actions

📁 Project Structure

.
├── code/               # SafeChat logic for generating chatbot files
├── data/               # All provided and intermediate data files
├── doc/                # Documentation and design assets
└── rasa_template/      # Template directory for RASA Open Source

🤝 Contributing

1. Fork the repository
2. Create your feature branch (git checkout -b feature/AmazingFeature)
3. Commit your changes (git commit -m 'Add some AmazingFeature')
4. Push to the branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

🛠️ Usage Examples

The following table highlights different use cases for SafeChat:

Domain	Description	Details
Election Information	Safe chatbot behavior for promoting voter engagement and participation	Disseminating Authentic Public Messages using Chatbots - A Case Study with ElectionBot-SC[1] On Safe and Usable Chatbots for Promoting Voter Participation[2]
Financial Advice	LLMs for fairness and efficacy in decision-makin	LLMs for Financial Advisement: A Fairness and Efficacy Study in Personal Decision Making[3]

📝 Citation

If you use SafeChat in your work, please cite the following publication:

@inproceedings{safechat-arch-github,
  title={SafeChat: A Framework for Building Trustworthy Collaborative Assistants (Github)},
  author={Muppasani, Bharath and Lakkaraju, Kausik and Gupta, Nitin and Nagpal, Vansh and Jones, Sara Rae and Srivastava, Biplav},
  booktitle={\url{https://github.com/ai4society/trustworthy-chatbot}},
  year={2024}
}

📚 Publications

1. Disseminating Authentic Public Messages using Chatbots - A Case Study with ElectionBot-SC to Understand and Compare Chatbot Behavior for Safe Election Information in South Carolina; Nitin Gupta, Vansh Nagpal, Bharath Muppasani, Kausik Lakkaraju, Sara Jones, Biplav Srivastava; Workshop on AI for Public Missions (AIPM) at Thirty-Ninth AAAI Conference on Artificial Intelligence (AAAI-25), Philadelphia, USA, Feb 2025

2. On Safe and Usable Chatbots for Promoting Voter Participation; Bharath Muppasani, Vishal Pallagani, Kausik Lakkaraju, Shuge Lei, Biplav Srivastava, Brett Robertson, Andrea Hickerson, Vignesh Narayanan; AAAI AI Magazine 2023

3. LLMs for Financial Advisement: A Fairness and Efficacy Study in Personal Decision Making; Kausik Lakkaraju, Sara Rae Jones, Sai Krishna Revanth Vuruma, Vishal Pallagani, Bharath C Muppasani and Biplav Srivastava; 4th ACM International Conference on AI in Finance: ICAIF'23, New York, 2023