Project Overview

Dynamic Query Resolution System

Overview

Welcome to the Dynamic Query Resolution System – a robust and scalable solution designed to handle text conversations in all 22 major languages of India. This project integrates cutting-edge technologies, including AI4Bharat ASR & Translational models, FAISS vector database, Gemini API, and various HuggingFace models to deliver efficient and context-aware responses. 🌐✨

Key Features

1. Speed Modes

   • 🚀 Superfast: Provides responses in 40ms using semantic search.
   • 🚄 Fast/Normal: Utilizes semantic search and Large Language Models for context-aware responses.

2. GPU Utilization

   • 🎮 All models have undergone quantization and are integrated onto a singular P100 GPU for streamlined and efficient GPU utilization.

3. Personalization

   • 🎨 Recommendations based on Contextual Bandits, Collaborative Filtering, Machine Learning, and geo-specific preferences enhance user personalization.

4. Text and Intention Classification

   • 🤖 A BERT and Keras based model serves as a multi-class text and intention classifier, augmenting the recommendation system.

5. Dataset Enhancement

   • 📊 Fine-tuning and Reinforcement Learning with Transformer Reinforcement Library by HuggingFace on custom-created Bhuvan datasets enhance model performance.

6. User Interface

   • 💻 Developed a React-based Web Application with a user-friendly interface capable of receiving multi-modal input, including both textual and voice inputs.

7. Backend Implementation

   • 🚀 FastAPI (Python) serves as the hosting platform for models, inferences, and API functionalities.

8. Deployment

   • 🚀 Backend deployed on Kaggle, harnessing the computational power of GPU resources.
   • 🔐 Secure private tunnel established between the frontend and backend using Ngrok.

9. Database Integration

   • 📂 Firebase, a non-relational database, stores crucial project data, including user conversation history, queries, and preferences. This database plays a key role in enabling optimal context derivation during each session.

Getting Started

Clone the repository:

git clone https://github.com/IISF-SIF/Narad

Demonstration Video

Watch the demonstration video on Google Drive 🎥: https://drive.google.com/drive/folders/11ODoqNLnvPUhRi3eZ50uVys6g3HXM468

Contributors

Arav Jain: Developed the multilingual, multi-speed RAG pipeline for delivering efficient, context-aware responses.

Vatsal Jha: Implemented the user-specific recommender system, finetuned and detoxified the models.

Ayushman Kar: Led backend implementation, database integration, and deployment.

Dilshad Sukheswala: Managed backend development and user interface design.