README.md

metadata

title: Status Law Gbot
emoji: 💬
colorFrom: yellow
colorTo: purple
sdk: gradio
sdk_version: 5.23.0
app_file: app.py
pinned: false

Status Law Assistant

An intelligent chatbot based on Hugging Face and LangChain for legal consultations using information from the Status Law company website.

📝 Description

Status Law Assistant is a smart chatbot that answers user questions about Status Law company's legal services. The bot uses RAG (Retrieval-Augmented Generation) technology to find relevant information in a knowledge base created from the official website content.

✨ Key Features

• Knowledge Base Management:
  • Dynamic URL management for knowledge base sources
  • Ability to add custom URLs for information extraction
  • Selective source inclusion/exclusion
  • Two modes of knowledge base updates:
    • Update Mode: Adds new information while preserving existing knowledge
    • Rebuild Mode: Complete recreation of knowledge base from selected sources
  • Real-time status tracking for knowledge base operations
  • Automatic metadata management and versioning
• Automatic creation and updating of knowledge base from status.law website content
• Intelligent information retrieval for query responses
• Context-aware response generation
• Advanced multilingual support:
  • Automatic language detection
  • Native language response generation
  • Built-in translation system with fallback mechanism
  • Support for 30+ languages
• Customizable text generation parameters
• Model switching system with automatic fallback
• Fine-tuning capabilities based on chat history
• Multiple model support:
  • Zephyr 7B: Enhanced performance and response quality
  • TinyLlama 1.1B Chat: Lightweight model for resource-constrained environments
  • Neural Mistral 7B: Superior reasoning and instruction following capabilities
  • Mixtral 8x7B: Advanced mixture-of-experts architecture

🚀 Technologies

• LangChain: Query processing chains and knowledge base management
• Hugging Face: Language model access and hosting
• FAISS: Efficient vector search
• Gradio: User interface creation
• BeautifulSoup: Web page information extraction
• PEFT: Efficient fine-tuning using LoRA
• SentencePiece: Tokenization

🏗️ Project Structure

status-law-gbot/
├── app.py                 # Main application file
├── requirements.txt       # Project dependencies
├── config/               # Configuration files
│   ├── settings.py       # Application and model settings
│   └── constants.py      # Constants and default values
├── src/                  # Source code
│   ├── analytics/        # Analytics module
│   │   └── chat_analyzer.py
│   ├── knowledge_base/   # Knowledge base management
│   │   ├── loader.py
│   │   └── vector_store.py
│   └── training/         # Training module
│       ├── fine_tuner.py
│       └── model_manager.py
└── dataset/             # HuggingFace dataset structure
    ├── annotations/     # Conversation annotations
    ├── chat_history/    # Chat logs and conversations
    ├── fine_tuned_models/ # Fine-tuned model storage
    ├── preferences/     # User preferences
    ├── training_data/   # Processed training data
    ├── training_logs/   # Training process logs
    └── vector_store/    # FAISS vector storage

💾 Data Storage

Dataset Organization

• annotations/: Conversation quality metrics and annotations
• chat_history/: JSON files containing chat conversations
• fine_tuned_models/: Storage for LoRA adapters and model checkpoints
• preferences/: User preferences and settings
• training_data/: Processed data ready for model training
• training_logs/: Detailed training process logs
• vector_store/: FAISS indexes for semantic search

🛠️ Setup

1. Clone the repository:

git clone https://github.com/PtOlga/status-law-gbot.git
cd status-law-gbot

2. Install dependencies:

pip install -r requirements.txt

3. Set up environment variables:

cp .env.example .env
# Edit .env with your configuration, including HUGGINGFACE_TOKEN

4. Run the application:

python app.py

🔧 Model Fine-tuning

To fine-tune the model on your chat history:

from src.training.fine_tuner import finetune_from_chat_history

success, message = finetune_from_chat_history(epochs=3)
print(message)

The fine-tuning process uses LoRA (Low-Rank Adaptation) for efficient training with minimal resource requirements.

🔄 Model Switching

The application supports multiple models with automatic fallback:

• Zephyr 7B: Enhanced performance and response quality
• TinyLlama 1.1B Chat: Lightweight model for resource-constrained environments
• Neural Mistral 7B: Superior reasoning and instruction following capabilities
• Mixtral 8x7B: Advanced mixture-of-experts architecture

Models can be switched dynamically through the interface or programmatically:

from src.training.model_manager import switch_to_model

switch_to_model("zephyr-7b")  # or "tinyllama-1.1b", "neural-mistral-7b", "mixtral-8x7b"

🔄 Knowledge Base Management

The application provides a flexible interface for managing knowledge sources:

1. Source Management:

   • View and edit the list of source URLs
   • Enable/disable specific sources
   • Add custom URLs for information extraction
   • Monitor source status and availability

2. Update Operations:

   • Update Knowledge Base: Incrementally add new information while preserving existing knowledge
   • Rebuild Knowledge Base: Completely recreate the knowledge base using only selected sources
   • Real-time operation status tracking
   • Automatic backup of previous versions

3. Usage:

   # Add new URL to knowledge base
   sources_df.append({"URL": "https://example.com", "Include": True, "Status": "Ready"})
   
   # Update knowledge base with selected sources
   update_kb_with_selected(sources_df)
   
   # Rebuild knowledge base from scratch
   rebuild_kb_with_selected(sources_df)
   

All changes to the knowledge base are automatically synchronized with the Hugging Face dataset, ensuring data persistence and version control.

🔗 Related Links

• Status Law Website
• Status Law Assistant on Hugging Face

📝 License

Public repository for Status Law Assistant.