Multi-Agent AI Assistant 🚀

This project is a Streamlit-based AI Assistant that allows users to train, fine-tune, and chat with AI models using OpenAI's API.

📌 Features

• Upload and Process Data (PDF, CSV, JSON)
• Preprocess Data for fine-tuning
• Evaluate Model for model evaluation tool
• Fine-tune AI Models with OpenAI's API
• Multi-Agent AI Chat with customizable prompts
• Restricted Words Filtering to control AI responses
• Vision-Language Model (VLM) for AI-powered image analysis 🖼️

🏗️ Project Structure

your_project/
│── main.py                    # Main entry point for Streamlit app
│── upload_and_train/
│   ├── upload_an_Train.py      # Handles file upload and preprocessing
│── fine_tune/
│   ├── fine_tune.py            # Fine-tuning OpenAI models
│── evaluate/
│   ├── evaluate_model.py       # Evaulate OpenAI models to the new trained model
│── chat/
│   ├── chat_with_model.py      # AI chatbot interactions
│── vlm/
│   ├── text_vlm.py             # AI Vision AI agent
│── utils/
│   ├── __init__.py             # Marks `utils/` as a package
│   ├── utils.py                # Contains helper functions
│── restricted_words.json       # Stores restricted words & model list (optional)
│── requirements.txt            # Python dependencies
│── README.md                   # Project documentation

📦 Installation

1️⃣ Clone the Repository

git clone https://github.com/TimGoebel/Multi_Agent.git
cd Multi_Agent

2️⃣ Set Up a Virtual Environment (Recommended)

python -m venv venv
source venv/bin/activate  # macOS/Linux
venv\Scripts\activate     # Windows

3️⃣ Install Dependencies

pip install -r requirements.txt
python -m spacy download en_core_web_sm

🚀 Running the App

Run the Streamlit app with:

streamlit run main.py

or use run.bat (for Windows users):

@echo off
python -m streamlit run main.py
pause  # Remove this if you don't want the CMD screen to stay open for troubleshooting

---

🛠️ Data Processing Pipeline

This project prepares and processes textual data from PDF, CSV, and JSON files for fine-tuning GPT models. It ensures high-quality, structured training data by applying advanced text cleaning, chunking, and formatting. The pipeline guarantees a minimum of 10 well-structured training examples, which is required for OpenAI fine-tuning.

1️⃣ Upload & Extract Data

• Supported Formats: PDF, CSV, JSON
• File Upload: Users can upload training data via Streamlit’s file uploader.
• Extraction Process:
  • CSV → Loaded into a DataFrame.
  • JSON → Parsed as structured text.
  • PDF → Text is extracted page by page using PyPDF2.

---

2️⃣ Advanced Text Cleaning

To improve model quality, raw text is cleaned using the following techniques:

• Unicode Normalization: Fixes encoding inconsistencies (unicodedata.normalize("NFKC", text)).
• HTML Tag Removal: Removes any embedded HTML content (BeautifulSoup).
• Hyphenation Fixes: Merges words split across lines (cross-\nword → crossword).
• Extra Whitespace Removal: Normalizes spacing (re.sub(r"\s+", " ", text)).
• Smart Quotes & Symbol Conversion: Converts to standard ASCII quotes and dashes (“” → ", — → -).
• Spell Checking & Correction: Fixes common spelling errors (SpellChecker).

---

3️⃣ Text Chunking & Structuring

To prevent mid-sentence breaks, the extracted text is split into meaningful units using spaCy:

• NLP Sentence-Based Chunking: Instead of breaking text arbitrarily at chunk_size, it ensures semantic coherence.
• Adaptive Chunking: If a chunk exceeds 1000 characters, it is split into smaller, logical parts.

The UI supports flexible chunking and measurement settings:

• Chunking Method: Choose between Sentence-level or Paragraph-level segmentation.
• Measurement Type: Select how chunk length is measured — by Character count or Token count.

---

4️⃣ Formatting for GPT-3 Fine-Tuning

Each data sample is formatted as a conversational exchange:

{
  "messages": [
    {"role": "system", "content": "You are an AI trained on specialized data."},
    {"role": "user", "content": "How does Mobile Lock protect authentication?"},
    {"role": "assistant", "content": "Mobile Lock prevents unauthorized access by restricting authentication when threats are detected."}
  ]
}

• Ensures well-structured conversations.
• Prevents empty "user" messages.
• Pairs each prompt with a valid assistant response.

---

5️⃣ Ensuring Minimum of 10 Examples

• If the dataset contains fewer than 10 training pairs, additional examples are generated automatically.
• Large text blocks are split into multiple prompts, ensuring a diverse dataset.
• Default responses are used where assistant completions are missing.

---

6️⃣ Exporting to JSONL for Fine-Tuning

The final dataset is saved in JSONL format, ready for OpenAI fine-tuning:

training_data.jsonl

Each entry is saved on a new line to comply with OpenAI’s fine-tuning requirements.

---

📌 How to Use

1️⃣ Upload & Train

• Upload a PDF, CSV, or JSON file.
• The data is chunked and preprocessed for fine-tuning.

2️⃣ Fine-Tune Model

• Train a new model using OpenAI’s fine-tuning API.
• Monitors training progress until completion.

3️⃣ Save & Train

• Click "Save Training Data" to export training_data.jsonl.
• Use OpenAI’s fine-tuning API to train the model.

---

🚀 Fine-Tuning Instructions

1️⃣ Prepare the Training Data

Ensure the dataset is ready in JSONL format:

openai tools fine_tunes.prepare_data -f training_data.jsonl

2️⃣ Train the Model

Fine-tune using OpenAI’s API:

openai api fine_tunes.create -t training_data.jsonl -m gpt-3.5-turbo

3️⃣ Deploy and Test

Use the fine-tuned model in your AI assistant.

---

🏗️ Future Enhancements

• Add local model fine-tuning support.
• add embedings
• Expand AI assistant multi-modal capabilities.
• Integrate vector search for improved retrieval.

📜 License

This project is licensed under the MIT License.

---

🔗 Connect with Me

• GitHub: Timothy Goebel
• LinkedIn: Timothy Goebel