README.md

metadata

title: Pathshala AI
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 4.44.1
app_file: app.py
python_version: 3.11
pinned: false

Pathshala AI

Pathshala AI is a bilingual AI tutor demo for rural primary students in Nepal.

This Hugging Face Space supports:

• Uploading a text-based PDF textbook directly in the Space
• Asking questions in English, Nepali, or romanized Nepali
• Retrieving relevant textbook portions from the uploaded PDF
• Showing a simple English answer and Nepali explanation
• Generating Nepali quiz questions
• Basic quiz grading

For the full web-app workflow, deploy the FastAPI backend separately and add a Space variable named BACKEND_URL.

Without BACKEND_URL, the Space can still run the same style of workflow locally. Add these Space secrets/variables to match the web app more closely:

• LLM_PROVIDER=qwen, LLM_BASE_URL, LLM_API_KEY, LLM_MODEL for the Qwen/vLLM tutor
• LLM_PROVIDER=gemini, GEMINI_API_KEY, GEMINI_MODEL for the Gemini tutor
• TRANSLATION_PROVIDER=gemini, GEMINI_API_KEY, GEMINI_MODEL for Nepali adaptation and romanized question normalization
• TRANSLATION_PROVIDER=openai, OPENAI_API_KEY, OPENAI_MODEL if you want to use OpenAI for Nepali adaptation instead
• OCR_PROVIDER=gemini, OCR_MAX_PAGES=5 for scanned or custom-font PDFs
• DEPLOYMENT_WARNING_ENABLED=true, DEPLOYMENT_WARNING_MESSAGE=... to show a visible warning when hosted credits or model services are unavailable

Use LLM_MODEL=Qwen/Qwen3-4B-Instruct-2507 to match the project default unless your vLLM endpoint exposes a different model name. Use GEMINI_MODEL=gemini-2.5-flash for stable Gemini API access, or gemini-3-flash-preview only if that preview model is enabled for your project.