📄 AskMyPDF

AskMyPDF is a Flask-based PDF Question Answering (RAG) web application that allows users to upload a PDF and ask questions about its content. The system answers only from the uploaded PDF, using semantic search and a local LLM, with clear source page references.

---

🚀 Features

• 📤 Upload a PDF file
• 🧠 Semantic search over PDF content (vector embeddings)
• 💬 Chat-style interface (Instagram / WhatsApp style)
• 📚 Answers strictly grounded in the PDF (no hallucination)
• 📄 Source page references for each answer
• 🔄 Reset session (PDF + chat)
• ⚡ No page reloads (AJAX / Fetch API)
• 🔐 Session-based chat memory
• 🧠 Local LLM inference (no OpenAI key required)

---

🏗️ Architecture Overview

This project follows a Retrieval-Augmented Generation (RAG) architecture:

PDF → Text Extraction → Chunking → Embeddings → Vector DB (Chroma)
                                  ↓
User Question → Semantic Search → Relevant Chunks → LLM → Answer

Tech Stack

Layer	Technology
Backend	Flask (Python)
Frontend	HTML, CSS, JavaScript
Templating	Jinja2
Vector DB	ChromaDB
Embeddings	Sentence-Transformers (MiniLM)
LLM	Ollama (Gemma 3)
PDF Parsing	PyMuPDF
Session Storage	Flask Sessions

---

📂 Project Structure

askmypdf/
├── app.py                # Flask application
├── rag.py                # RAG pipeline (index + retrieval + LLM)
├── uploads/              # Uploaded PDF files
├── chroma_store/         # Persistent vector database
├── templates/
│   ├── base.html
│   └── index.html
├── static/
│   ├── style.css
│   └── app.js
├── config/
│   └── .env
├── screenshots/ # Application screenshots
│    └──01-home-empty.png
│    └──02-upload-pdf.png
│    └──03-chat-question.png
│    └──04-chat-answer.png
│    └──05-multiple-qa-demo.png
│    └──06-qa-proof-mywalletpro-tech-page2.png
│    └──07-negative-queries-not-in-pdf.png
│    └──08-reset-state.png
├── .gitignore
├── LICENSE
├── requirements.txt
└── README.md

---

🧠 How It Works (Step by Step)

1. User uploads a PDF
2. PDF text is extracted using PyMuPDF
3. Text is split into overlapping chunks
4. Each chunk is converted into a vector embedding
5. Vectors are stored in ChromaDB
6. User asks a question
7. Most relevant chunks are retrieved using semantic similarity
8. LLM verifies if the answer exists in the PDF
9. LLM generates a final answer only from the PDF
10. Source page numbers are returned

---

🛡️ Anti-Hallucination Strategy

• LLM is not allowed to use outside knowledge
• A verification step checks whether the answer exists in the retrieved text
• If not found, the system replies:

Answer is not in this PDF.

---

🖥️ User Interface

• Chat bubbles (user → right, assistant → left)
• Smooth scrolling chat
• Fixed bottom input bar
• 🎤 Voice input (speech-to-text)
• Status messages (uploading, thinking, done)
• PDF name badge in navbar
• Reset button clears PDF + chat without reload

---

⚙️ Installation & Setup

1️⃣ Clone the repository

git clone https://github.com/khaledelsayed2003/askmypdf.git
cd askmypdf

2️⃣ Create a virtual environment

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

3️⃣ Install dependencies

pip install -r requirements.txt

4️⃣ Install Ollama

Download from: 👉 https://ollama.com

Pull the model:

ollama pull gemma3:4b

5️⃣ Environment variables

Create config/.env:

FLASK_SECRET_KEY=your-secret-key

6️⃣ Run the app

python app.py

Open in browser:

http://127.0.0.1:5000

---

🧪 Example Usage

1. Upload a PDF (CV, paper, report, book)

2. Ask:

   • “What programming languages are listed?”
   • “What is the project objective?”
   • “Who is the author?”

3. Get:

   • A concise answer
   • Exact source page(s)

---

🔄 Reset Behavior

• Clears:

  • Uploaded PDF
  • Vector collection
  • Chat history

• UI resets instantly (no reload)

---

📸 Screenshots & Demo Walkthrough

This section demonstrates the full workflow of AskMyPDF, from uploading a document to asking questions, validating answers, handling unsupported queries, and resetting the session.

1. Initial Home Screen (No PDF Uploaded)

When the application starts, no document is loaded. The user is prompted to upload a PDF before asking any questions.

2. Uploading a PDF

The user uploads a PDF file, which is:

• Saved locally
• Indexed into a vector database (ChromaDB)
• Prepared for semantic search

3. Asking a Question

Once a PDF is uploaded, the user can ask questions using a chat-style interface. User messages appear on the right, similar to modern messaging apps.

4. Answer Generated from PDF

The assistant retrieves relevant content from the PDF and generates an answer. Each answer includes the exact source page number(s) + it might give also the source page for the second highest retrieved vector.

5. Multiple Questions in a Single Session

Users can ask multiple questions without reloading the page. The full conversation history is preserved until reset.

6. Source Page Verification (Proof)

This example demonstrates that the answer is:

• Correct
• Present in the PDF
• Referenced with the correct page number

Example shown: MyWallet Pro technologies on page 2.

7. Hallucination Prevention (Unsupported Questions)

If a question is not answered explicitly in the PDF, the system responds with:

"Answer is not in this PDF."

This ensures the assistant never fabricates information.

8. Resetting the Session

The reset button:

• Clears the uploaded PDF
• Deletes the vector index
• Clears chat history
• Returns the app to its initial state

---

🧾 Limitations

• Works with text-based PDFs (not scanned images)
• Large PDFs may take longer to index
• Requires local LLM (Ollama) to be running

---

🛠️ Future Improvements

• 📑 Multi-PDF support
• 🔍 Highlight answer text inside PDF
• 📊 Confidence scoring
• 🌐 Deployment (Docker / Cloud)

---

📜 License

This project is licensed under the MIT License. You are free to use, modify, and distribute it.

---

👤 Author

Khaled Elsayed

• 📧 Email: khaled.elsayed2206@gmail.com
• 🧠 Built for learning, research, and real-world RAG systems

---