Skip to content

ankitdutta428/smart-chatbot

Repository files navigation

CARE: Customer Assistance and Response Engine

Overview

CARE (Customer Assistance and Response Engine) is an advanced AI-powered chatbot designed to provide efficient, empathetic, and domain-specific support across various industries, including:

  • Banking
  • Healthcare
  • E-commerce
  • Telecommunications

The chatbot leverages Phi 3.5 Mini Instruct, a compact, instruction-tuned language model, optimized for efficient fine-tuning and deployment. CARE aims to automate customer interactions, analyze sentiment, and escalate unresolved issues seamlessly.


Features

  • Domain-Specific Responses: Customized support for diverse industries like healthcare, banking, telecommunications, and e-commerce.
  • Sentiment Analysis: Provides empathetic responses by analyzing user sentiment.
  • Efficient Performance: Built using a compact model optimized for faster inference and lower memory requirements.
  • Extensible Framework: Easily expandable to support additional industries or integrate with new systems.

Execution Framework

The project employs the following workflow:

  1. Input: User prompts specific to the industry (e.g., banking, healthcare).
  2. Processing:
    • Sentiment analysis for empathetic responses.
    • Fine-tuned language model tailored for the respective domain.
  3. Output: Generates accurate, domain-specific responses and escalates issues if needed.

Phi 3.5 Mini Instruct ensures:

  • Instruction-tuned training for contextual understanding.
  • Transformer architecture for balanced performance and efficiency.

Datasets Used

CARE is trained on domain-specific datasets to enhance accuracy:

Custom datasets were also utilized for enhanced domain adaptation.


Methodology

  • Model Architecture: Fine-tuning and QLoRA techniques were used to optimize Phi 3.5 Mini Instruct.
  • Training Pipeline:
    • Data preprocessing: Addressed dataset skewness using oversampling and undersampling techniques.
    • Model merging: Combined fine-tuned domain-specific models into a unified framework.
  • Final Model: Phi-3.5-mini-instruct-Combined-Model
  • Testing: Testing Notebook

Possible Improvements

Future enhancements for CARE could include:

  1. Multi-Language Support: Expand capabilities to handle queries in multiple languages.
  2. Voice Integration: Real-time voice-to-text and text-to-voice for improved accessibility.
  3. Reinforcement Learning with Human Feedback (RLHF): Continuously improve chatbot accuracy using RLHF and Retrieval-Augmented Generation (RAG).
  4. Knowledge Base Expansion: Regularly update databases for better accuracy and relevance.
  5. API Deployment: Make the chatbot available as an API to expand accessibility and impact.

Conclusion

CARE represents a step forward in smart customer service, providing solutions tailored to diverse industries. By leveraging advanced NLP techniques and domain-specific training, CARE enhances customer satisfaction and streamlines issue resolution.


Authors

  • Ankit Dutta
  • Nabarup Ghosh
  • Ankush Chatterjee
  • Abhishek Das

Special thanks to Genpact and Kshitij for providing the opportunity to develop this impactful project.


License

This project is licensed under the AGPL-3.0 License. See the LICENSE file for details.


About

This repo contains the code for making the smart chatbot made using finetuned large language model dpeloyment

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors