RAG Powered Question-Answering System

This project is a Retrieval-Augmented Generation (RAG) based Question-Answering system. It uses a language model to answer questions based on the content of a provided PDF or text file. The system is built with a FastAPI backend and a Next.js frontend.

Features

• File Upload: Upload PDF or text files to be used as the knowledge base.
• Question-Answering: Ask questions about the content of the uploaded file and get detailed answers.
• Dynamic UI: The user interface is built with Next.js + React and provides a ChatGPT-style experience.
• Markdown, LaTeX & Code Rendering: Responses render Markdown, KaTeX equations, tables, and syntax-highlighted code blocks.
• Responsive Layout: Optimised for both desktop and mobile viewing.

Installation

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.

Prerequisites

• Python 3.9+
• Pip (or Conda)
• Node.js 18+

Setup Using uv

1. Clone the repository:

   git clone https://github.com/Ghost-141/AskPDF-AI.git
   cd AskPDF-AI

2. Create a virtual environment (recommended):

   uv sync
   .\.venv\Scripts\activate   # Windows
   source .venv/bin/activate  # macOS / Linux

3. Install the dependencies:

   pip install -r requirements.txt

Setup Using Conda

1. Clone the repository:

   git clone https://github.com/Ghost-141/AskPDF-AI.git
   cd AskPDF-AI

2. Create a Conda environment:

   conda create --name pdf_qa python=3.9.23
   conda activate pdf_qa

3. Install the dependencies:

   pip install -r requirements.txt

4. Install the frontend dependencies:

   cd frontend
   npm install

GPU Support (NVIDIA)

For GPU acceleration, you need to install PyTorch with CUDA support. Make sure you have the correct NVIDIA drivers and CUDA Toolkit version installed.

1. Check your NVIDIA driver and CUDA version: You can check your NVIDIA driver version by running nvidia-smi in your terminal. This will also show the highest version of CUDA that is supported.

2. Install PyTorch with CUDA: Visit the PyTorch website to find the correct command for your specific CUDA version. For example, to install PyTorch with CUDA 12.6, you would run:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126

   Note: Using a version of PyTorch with CUDA support is essential for running the model on an NVIDIA GPU. If you do not have a compatible GPU, the embedding model will run on the CPU, which will be significantly slower while processing the pdf.

Document parsing (unstructured)

File extraction uses unstructured to perform layout-aware PDF partitioning. This gives the retriever much cleaner chunks: heading-aware boundaries, typed metadata (Title, NarrativeText, ListItem, Table, …), and tables rendered as Markdown so the LLM and the frontend's react-markdown can both consume them without alignment loss.

The default partitioning strategy is "fast" — pure-Python, no OCR, works in a stock install. If you want layout-aware table detection (and you've installed the OCR extras), set FileProcessor(..., strategy="hi_res") in backend/api/files.py. The hi_res strategy is slower at ingestion time but yields more accurate table extraction.

If unstructured raises on a problematic file (rare, usually due to a missing OCR dependency or an exotic PDF), backend/services/file_processor.py automatically falls back to the previous PyMuPDFLoader / TextLoader path — you will see a unstructured extraction failed ... falling back log line.

OCR for scanned PDFs is not enabled by default; install unstructured_pytesseract and switch to strategy="hi_res" if you need it.

Configuration

Create a .env file in the root of the project and add your keys:

GROQ_API_KEY=your-groq-api-key
GROQ_MODEL_NAME=openai/gpt-oss-120b   # optional override

Uploads are stored in data/raw and the vector database persists in data/vector_db (created automatically).

Running the Application

Start the backend and frontend in separate terminals.

Backend (FastAPI)

Run via the uv-managed virtual environment (so the project's pinned langchain/unstructured/uvicorn versions are used, not whatever happens to be on PATH from another Python install):

# from the repo root
uv run --python 3.11 python -m uvicorn backend.main:app --reload

Avoid uv run rav run backend — rav is a third-party wrapper that shells out to its own bundled uvicorn (often from a different Python install), which will fail with ModuleNotFoundError: No module named 'langchain_core' even when the project's venv has everything installed. Use the python -m uvicorn form above; it always uses the venv's uvicorn.

The API is available at http://localhost:8000.

Frontend (Next.js)

cd frontend
npm install      # first time only
npm run dev

The UI is available at http://localhost:3000.

Tip: Set NEXT_PUBLIC_API_URL inside frontend/.env.local if your API is not running on the default host/port.

Project Structure (relevant parts)

• backend/main.py: FastAPI app factory and CORS setup.
• backend/api/: API routers for files and QA (/upload-file, /process-file, /ask).
• backend/services/: File processing, vector store management, QA pipeline, uploads.
• backend/data/: Runtime data; data/raw for uploads, data/vector_db for Chroma persistence.
• frontend/: Next.js client.

Usage

1. Upload a file: Use the upload card in the UI to select a PDF or text file.
2. Process the file: Click "Upload & Process" to push the file to the backend and populate the vector store.
3. Ask a question: Type your question in the chat composer. Responses appear in a ChatGPT-style transcript with full formatting.

API Documentation

The FastAPI backend provides the following endpoints:

• POST /upload-file: Accepts a multipart file upload and stores it in data/raw.
  • Form Field: file (UploadFile, required)
  • Success Response: {"filename": "<stored-filename>"}
• POST /process-file: Processes a previously uploaded file.
  • Query Parameter: filename (string, required)
  • Success Response: {"message": "File '<filename>' processed...", "num_docs": <number>}
  • Error Response: {"detail": "File '<filename>' not found..."}
• POST /ask: Asks a question to the model.
  • Request Body: {"query": "<your-question>"}
  • Success Response: {"answer": "<model-answer>"}
  • Error Response: {"detail": "QA pipeline not initialized..."}

Libraries & Frameworks Used

• FastAPI
• Next.js
• React
• LangChain
• ChromaDB
• PyPDF / PyMuPDF

Upcoming Features

• Support for Bangla Languge
• Process Images and Complex Pdf
• Support for different documents
• Local llm support