A Streamlit-based chatbot that combines document retrieval with language generation capabilities. This chatbot can translate input, retrieve relevant documents from a corpus, and generate contextual responses.

• Multi-language Support: Automatically translates user queries using Google Translator
• Semantic Search: Retrieves relevant documents using sentence embeddings and FAISS
• Conversational Memory: Maintains conversation history for context-aware responses
• LLM-Powered Responses: Generates human-like responses using Llama 3.1 (8B parameter model)
• Interactive Web Interface: Built with Streamlit for easy interaction

The chatbot consists of three main components:

1. Retriever: Uses SentenceTransformer with the 'all-MiniLM-L6-v2' model to create embeddings and FAISS for efficient similarity search
2. Generator: Implements Meta's Llama 3.1 (8B) model with 4-bit quantization for memory efficiency
3. Translator: Incorporates Google Translator for handling non-English queries

• Python 3.8+
• PyTorch
• CUDA-compatible GPU (recommended)
• Hugging Face account for model access

# Clone the repository
git clone https://github.com/yourusername/conversational-retrieval-chatbot.git
cd conversational-retrieval-chatbot

# Create a virtual environment
python -m venv venv
source venv/bin/activate  # On Windows, use: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Log in to Hugging Face (required to access Llama 3.1)
huggingface-cli login

Create a requirements.txt file with the following dependencies:

streamlit
faiss-cpu  # Use faiss-gpu if using GPU
torch
sentence-transformers
transformers
deep-translator
huggingface-hub
accelerate
bitsandbytes

1. Run the Streamlit app:

   streamlit run main.py

2. Access the app in your browser (typically at http://localhost:8501)

3. Enter your corpus in the sidebar (one document per line)

4. Start chatting in the main interface

To use a different sentence embedding model, modify:

self.retriever_model = SentenceTransformer('your-preferred-model')

Modify the generation settings in the generate_answer method:

outputs = self.generator_model.generate(
    **inputs,
    max_length=100,         # Adjust maximum response length
    temperature=0.5,        # Control randomness (higher = more random)
    top_p=0.9,              # Nucleus sampling parameter
    num_return_sequences=1, # Number of response candidates
    do_sample=False         # Set to True for more varied responses
)

Change the default translation languages:

def translate_text(self, text, source_lang='your-source-lang', target_lang='your-target-lang'):

• Requires significant GPU memory for the Llama 3.1 model, even with quantization
• Performance may vary based on the quality and relevance of the provided corpus
• Translation quality depends on Google Translator's capabilities

• Hugging Face for transformer models
• FAISS for efficient similarity search
• Sentence Transformers for text embeddings
• Streamlit for the web interface
• Meta AI for the Llama 3.1 model

Note, I have writing readme