Abrahao Labs AI Chatbot

An intelligent chatbot specialized in digital marketing for Abrahao Labs, built with Python, LangChain, and Hugging Face fine-tuned models. Features RAG (Retrieval Augmented Generation) for contextual responses from external knowledge bases.

Features

• 🤖 Automated Customer Service: Natural conversations about digital marketing services
• 🔍 RAG-Powered Responses: Retrieves context from PDF, CSV, and TXT documents
• 📚 Knowledge Base: Integrates information from multiple file formats
• 💼 Service Presentation: Explains available digital marketing services
• 💰 Personalized Recommendations: Suggests plans based on customer profiles
• 🔄 Context Retention: Maintains conversation history for coherent interactions
• 🎯 Lead Qualification: Identifies customer needs and qualifies leads automatically

Tech Stack

• Python 3.9+: Core programming language
• LangChain: Framework for LLM applications and RAG
• Hugging Face Transformers: Fine-tuned language model (distilgpt2)
• Flask: Web framework for the chat interface
• FAISS: Vector storage for efficient semantic search
• Sentence-Transformers: Semantic embeddings generation
• PyPDF2 & Pandas: Document processing (PDF, CSV)

Project Structure

├── app/
│   ├── __init__.py
│   ├── chatbot.py              # Main ChatBot class
│   ├── conversation_manager.py # Conversation state & lead qualification
│   ├── data_loader.py          # Document loading (PDF, CSV, TXT)
│   ├── prompt_manager.py       # LLM prompt templates
│   ├── vector_store.py         # FAISS vector storage
│   └── fine_tuning.py          # Model fine-tuning module
├── data/
│   ├── faq.csv                 # FAQ about digital marketing
│   ├── pricing_plans.csv       # Plans and pricing information
│   ├── lead_qualification_script.txt  # Lead qualification script
│   └── agency_services.pdf     # Detailed service descriptions
├── fine_tuned_model/           # Fine-tuned model (local)
├── vectorstore/                # FAISS vector database
├── templates/
│   └── index.html              # Web chat interface
├── static/
│   ├── logo.png                # Abrahao Labs logo
│   ├── script.js               # Frontend JavaScript
│   └── style.css               # CSS styles
├── app.py                      # Flask server
├── run_fine_tuning.py          # Fine-tuning script
├── .env                        # Environment variables
└── requirements.txt            # Dependencies

Installation

1. Clone the repository:

   git clone https://github.com/abrahao-dev/abrahao-labs-chatbot
   cd abrahao-labs-chatbot

2. Create and activate a virtual environment:

   python -m venv .venv
   source .venv/bin/activate  # On Windows: .venv\Scripts\activate

3. Install dependencies:

   pip install --upgrade pip
   pip install -r requirements.txt

4. Configure environment variables (optional for offline mode):

   echo "HUGGINGFACE_API_KEY=your_api_key" > .env

5. Get the fine-tuned model files:

   The fine-tuned model files are not included in the repository due to GitHub size limits.

   • Option A (Recommended): Download from Google Drive and extract to fine_tuned_model/
   • Option B: Run fine-tuning locally (may take several hours):

     python run_fine_tuning.py

6. Run the application:

   python app.py

7. Access the web interface at: http://localhost:8080

How It Works

RAG (Retrieval Augmented Generation)

The chatbot uses RAG to enhance responses with relevant information from:

• agency_services.pdf: Digital marketing service descriptions
• faq.csv: Frequently asked questions with answers
• pricing_plans.csv: Plan details and pricing
• lead_qualification_script.txt: Lead qualification workflow

Documents are processed, indexed, and stored in a FAISS vector database for efficient semantic search.

Fine-Tuning

The project fine-tunes a base model (distilgpt2) for digital marketing specialization:

1. Data Preparation: Training conversations are formatted appropriately
2. Tokenization: Text is converted to tokens
3. Training: Model is fine-tuned with specific text generation settings
4. Local Storage: Trained model is saved for offline use

Customization

• Prompts: Customize LLM instructions in app/prompt_manager.py
• Training Data: Edit files in data/ to update the knowledge base
• Fine-tuning: Run python run_fine_tuning.py to retrain with new data
• Interface: Modify templates/index.html and static/ files

System Requirements

• Python 3.9 or higher
• Minimum 4GB RAM
• At least 1GB free disk space
• Dual-core processor or better

Troubleshooting

Issue	Solution
Model loading error	Verify fine-tuned model files are in fine_tuned_model/
Dependency errors	Run pip install -r requirements.txt again
Slow performance	Reduce document count in knowledge base
Port in use	Change port in app.py (default: 8080)

Known Limitations

• Fine-tuning can be slow on machines without GPU
• Fine-tuned model may occasionally generate lower quality responses
• Short messages may be difficult to process for lead qualification
• Some edge cases in vector storage integration need refinement

Authors

Developed as an AI course project at Universidade São Francisco (USF).

• Matheus Abrahão Martins Alvares
• Cauê Becker Rossi
• Bruno Ricci Zanoni
• Giulio Passetti Figueiredo

License

MIT License - © 2026 Abrahao Labs