Intelligent Document Classification System

This project classifies uploaded PDF documents into business-oriented categories using a small NLP and machine learning pipeline. It combines a Streamlit frontend for file upload and preview, a FastAPI backend for inference, and a trained scikit-learn model for prediction.

Highlights

• Upload and preview PDF files in the Streamlit interface
• Extract document text with PyPDF2
• Clean and normalize text with NLTK-based preprocessing
• Transform text with TF-IDF vectorization
• Predict document categories with logistic regression
• Return a confidence score and confidence level for each classification

Architecture

• Architecture diagram
• Screenshots folder

Project Structure

document_classification_system/
|-- app/
|   |-- main.py
|   |-- train_model.py
|   |-- services/
|   |   `-- classifier.py
|   |-- trained_models/
|   |   |-- model.pkl
|   |   `-- vectorizer.pkl
|   `-- utils/
|       |-- pdf_extractor.py
|       `-- preprocessing.py
|-- dataset/
|   `-- documents.csv
|-- screenshots/
|-- temp_uploads/
|-- test_documents/
|-- requirements.txt
`-- streamlit_app.py

How It Works

1. A user uploads a PDF in the Streamlit app.
2. The frontend sends the file to the FastAPI classification endpoint.
3. The backend stores the uploaded file temporarily.
4. pdf_extractor.py reads the PDF and extracts plain text.
5. preprocessing.py normalizes the text and removes stopwords.
6. classifier.py loads the saved TF-IDF vectorizer and logistic regression model.
7. The backend returns the predicted category, confidence score, and confidence level.
8. The Streamlit UI displays the result and a progress bar.

Model Training

Training is handled by app/train_model.py. The script:

• loads dataset/documents.csv
• drops missing rows
• preprocesses the document text
• vectorizes text with TfidfVectorizer
• trains a LogisticRegression model
• saves the model artifacts to app/trained_models/

Expected Document Categories

Based on the dataset and sample files in this repository, the system is currently oriented around categories such as:

• invoice
• ShippingOrder
• purchase order style documents
• resume
• sales report

The exact classes available at runtime depend on the labels present in dataset/documents.csv.

Running the Project

1. Create a virtual environment

python -m venv .venv
.\.venv\Scripts\Activate.ps1

2. Install dependencies

pip install -r requirements.txt
pip install streamlit nltk requests

3. Train or refresh the model

python -m app.train_model

4. Start the backend API

uvicorn app.main:app --reload

5. Start the Streamlit frontend

streamlit run streamlit_app.py

The frontend and backend should be started in separate terminals. Keep your runtime-specific connection details in local configuration or code as needed, rather than publishing them in project documentation.

API Response Shape

The classification endpoint returns a response in this format:

{
  "filename": "sample.pdf",
  "predicted_category": "invoice",
  "confidence_score": "96.41%",
  "confidence_level": "High"
}

Notes

• Uploaded files are temporarily stored in temp_uploads/.
• The current implementation is focused on PDF classification.
• requirements.txt does not currently include every imported runtime package used by the app UI and preprocessing flow, so the extra install step above is intentional.