QueryPDF

QueryPDF is a local-first, terminal-based RAG tool that allows users to interact with PDF documents using natural language, entirely offline. It extracts text from user-specified pages, chunks it intelligently, embeds the content, and stores it in a local vector database. Then, using an LLM served by Ollama, it provides accurate, grounded responses to your queries — all without any internet connection.

🔍 Features

• Privacy: All processing happens on your machine, no data leaves your system.
• Local Caching: Caches both tokenizers and embedding models locally for offline use and faster future runs.
• Interactive Page Selection: Specify which pages to analyze at runtime.
• Graceful Exit Handling: Automatically cleans up ChromaDB collections on exit.
• Visual Progress Feedback: Shows embedding generation progress.
• Context-Aware Responses: Uses semantic search to find the most relevant chunks for your questions.
• Fully Customizable: Users can configure embedding model, LLMs, and chunk parameters via a simple JSON file.

🛠️ Technologies Used

Technology	Purpose
pypdf	Extract text from PDF files
transformers	Tokenization using pretrained models
langchain	Text chunking with recursive character splitter
sentence-transformers	Generate dense vector embeddings
ChromaDB	Store and query embeddings locally
Ollama	Run open-source LLMs locally

✅ Requirements

• Python 3.10+
• Ollama
  • Install Ollama and run:

    ollama run gemma3:1b

📦 Installation

Create a virtual environment using:

python -m venv <environment-name>

Activate the virtual environment:

./<environment-name>/scripts/activate

Install the dependencies:

pip install pypdf transformers sentence-transformers langchain chromadb ollama colorama yaspin

Or using requirements.txt:

pip install -r requirements.txt

📂 Project Structure

project/
├── config.json         # Configuration file
├── app.py              # Main script
├── sample.pdf          # Example PDF for testing
├── requirements.txt    # Package dependencies
└── README.md           # Documentation

⚙️ Configuration

The application uses a config.json file with the following parameters:

{
    "embedding_model": "intfloat/e5-small-v2",
    "ollama_model": "gemma3:1b",
    "chunk_size": 450,
    "chunk_overlap": 100
}

You can modify these settings to adjust:

• The embedding model used for semantic search (you can find embedding models on hugging face) .
• The Ollama LLM model used for response generation as per your requirements and resources (you can find ollama models on Ollama github page) .
• Make sure to install the required LLM model by running ollama run <model-name> in the terminal -- This installs the model locally .
• Chunk size and overlap for text splitting as per the model .

▶️ Usage

1. Run the script:

python app.py

2. When prompted, paste the full path to your PDF file:

Paste the PDF path: /path/to/your/document.pdf

3. Enter the page range you want to analyze:

Enter the page range (e.g., 15-25): 10-20

4. The application will process the PDF, extracting text, generating embeddings, and storing them in ChromaDB.

5. Enter your questions when prompted:

Chat with PDF: What is the main topic discussed in the document?

6. To exit the application, type:

Chat with PDF: /exit

⚙️ How It Works

1. PDF Extraction: Extracts text from specific pages based on user input (you specify the page range during runtime).

2. Tokenization & Chunking: Uses the specified tokenizer to split the text into chunks of configurable size (default: 450 tokens) with customizable overlap (default: 100 tokens).

3. Embeddings: Creates embeddings for each chunk using SentenceTransformers (default: intfloat/e5-small-v2).

4. Vector Storage: Stores the embeddings and original text chunks in a local ChromaDB collection.

5. Chat Interface: Accepts user queries, retrieves the top 3 most relevant chunks, and feeds them into an Ollama-served LLM (default: gemma3:1b) for response generation.

📁 Local Model Caching

To ensure full offline capability and improve performance, QueryPDF automatically saves required models locally:

• 🧠 Tokenizer: On first run, the Hugging Face tokenizer (transformers.AutoTokenizer) is downloaded and saved to a local_tokenizer/ directory. Subsequent runs will use this local tokenizer even without an internet connection.

• 🔤 Embedding Model: The embedding model (e.g., intfloat/e5-small-v2) is saved locally under local_e5_small_v2/ on the first run. If the model is already cached, it will be reused without redownloading.