UPSC Answer Script Evaluation UI

A Streamlit app to upload a scanned PDF of UPSC answer sheets, run OCR (Tesseract or Google Vision), parse question-answer pairs, and evaluate answers using a rubric. Works without LLM (heuristic baseline) or with OpenAI for improved parsing and scoring.

Features

• PDF to image rendering (PyMuPDF)
• OCR via:
  • Tesseract (local, free)
  • Google Cloud Vision (service account JSON)
• Q/A parsing:
  • Heuristic baseline (no LLM)
  • OpenAI model (e.g., gpt-4o-mini) for structured JSON extraction
• Evaluation:
  • Heuristic baseline scoring
  • OpenAI-based rubric scoring (Relevance, Accuracy, Depth, Structure, Language)
• JSON report download
• Toggle to skip the first two pages during OCR (useful to ignore cover/index pages)
• Optional factual grounding via web search to produce specific, cited comments

Setup (Windows / PowerShell)

1. Install Python 3.10+ and pip.

2. Create a virtual environment (optional but recommended):

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

3. Install Python dependencies:

pip install -r requirements.txt

4. Install Tesseract (for local OCR):

• Download Windows installer from: https://github.com/UB-Mannheim/tesseract/wiki
• After install, note the path to tesseract.exe (commonly C:\\Program Files\\Tesseract-OCR\\tesseract.exe).
• In the app sidebar, provide this path if Tesseract is not on PATH.

5. Optional: Configure Google Vision OCR

• Create a Google Cloud project and enable Vision API.
• Create a service account and download the JSON key (a .json file).
• In the app sidebar, choose "Google Vision" as the OCR provider and upload the JSON key file when prompted. The app stores it in a temporary file for this session and uses it for OCR.

6. Optional: Use OpenAI for parsing/scoring

• Set your OpenAI key in the sidebar, or export before running:

$env:OPENAI_API_KEY = "sk-..."

Run the app

streamlit run app.py

Then open the local URL shown (typically http://localhost:8501).

Usage Tips

• Start with Tesseract OCR for a quick baseline. If handwriting is poor, try Google Vision.
• If you don’t have an OpenAI key, the app still works with heuristic parsing and scoring.
• Adjust rubric weights in the sidebar based on the marking scheme.
• If your PDF's first two pages are covers or instructions, enable "Skip first 2 pages during OCR" in the sidebar. Rendering/preview still shows all pages, but OCR starts from page 3.
• Enable "Factual grounding via web search" when using OpenAI to make comments more specific and include citations. This uses DuckDuckGo search; no API key required.

Project Structure

• app.py – Streamlit UI
• core/ – Core logic package
  • evaluator.py – Evaluation and parsing logic
  • vlm.py – Vision Language Model integration
  • ocr.py – OCR handling (Tesseract/Google Vision)
  • pdf.py – PDF processing
  • common.py – Shared utilities
  • models.py – Data models
• requirements.txt – Python deps
• EvaluationPrototype.ipynb – Original prototype notebook

Troubleshooting

• PyMuPDF or Tesseract import errors: ensure pip install -r requirements.txt was successful.
• Tesseract not found: provide full path to tesseract.exe in the sidebar.
• Google Vision errors: verify the service account has Vision API access and that you uploaded a valid JSON key in the sidebar.
• OpenAI errors: check the model name (e.g., gpt-4o-mini) and key validity.
• Error cannot import name 'Sentinel' from 'typing_extensions': your environment has an outdated typing_extensions.
  • Fix:

    pip install -U typing_extensions
    # or ensure prod env uses the pinned version from requirements.txt
    pip install -r requirements.txt