Banking Customer Support System

This repository implements a chatbot system for banking customer support, leveraging Streamlit for the user interface, sentence transformers for intent classification, and Gemini LLM for response generation with contextual understanding.

Key Features:

• Speech Recognition: Handles user audio input using OpenAI Whisper for efficient transcription.
• Intent Classification: Classifies user queries into predefined intents based on sentence embeddings and cosine similarity.
• Contextual Response Generation: Utilizes Gemini LLM with chat history memory to provide informative and relevant responses.
• Text-to-Speech Output: Converts chatbot responses to audio for natural interaction.

Installation:

1. Clone this repository.

2. Create a virtual environment (recommended).

3. Install dependencies:

   pip install -r requirements.txt

4. Set the OpenAI API key in a .env file:

   OPENAI_API_KEY="your_openai_api_key"

Usage:

1. Run the application:

   streamlit run app.py

2. Interact with the chatbot through the Streamlit interface:

• Type your question or message in the text field.
• Alternatively, click the microphone icon and record your audio query.

Code Structure:

• app.py: Main application logic, managing Streamlit UI elements, handling user input, and coordinating chatbot interactions.
• chatbot.py: Defines the ChatBot class, responsible for response generation using LLMChain and intent classification based on sentence transformers.
• audio_utils.py: Provides functions for audio transcription (using OpenAI Whisper), text-to-speech conversion (using gTTS), and audio playback in the Streamlit app.

Technical Details:

Intent Classification:

Sentence embeddings are generated using a pre-trained sentence transformer model for user input and predefined intents. Cosine similarity is calculated between the user input embedding and all dataset embeddings to identify the most similar intent.

Chatbot Response Generation:

A ChatBot instance is created, incorporating an LLMChain object with a customized prompt template. The prompt template leverages the chat history memory to provide context-aware responses. The user's query and chat history are fed into the LLMChain to generate a response aligned with the classified intent.

Audio Processing:

OpenAI Whisper is used for transcribing speech input to text. gTTS library handles the text-to-speech conversion, saving speech output as a WAV file. Streamlit displays transcribed text and plays the generated speech file.