SmartAssessor

New Architecture (Firestore Queue)

Frontend → Backend (PC) → Firestore “jobs” → Colab worker (GPU LLM) → Firestore → Backend → Frontend.

The backend no longer runs a local LLM. Instead, it creates a job document in Firestore and waits for a Colab worker (running the GPU model) to process it and write back results.

Setup

• Firebase: create a project, enable Firestore (Native), and create a Service Account (Admin SDK JSON). Keep the JSON private.
• Env vars (PC + Colab):
  • FIREBASE_PROJECT_ID=smart-assessor-fdeaa
  • GOOGLE_APPLICATION_CREDENTIALS=<path to service-account.json>
• Python deps (PC + Colab):
  • pip install google-cloud-firestore google-auth

Backend (PC)

• Endpoints:
  • POST /upload/assessment — Accepts a PDF, extracts text, splits into chunks, enqueues a Firestore job, waits for result, and returns { evaluation: string }.
  • POST /grade-text — Smoke test; sends a single text as one job.
• CORS is configurable via ALLOWED_ORIGINS.

Run locally:

pip install -r backend/requirements.txt
uvicorn backend.app.main:app --reload --host 0.0.0.0 --port 8000

Colab Worker (GPU LLM)

• Notebook has two cells:
  1. Load model and define _llm.chat(text) returning a JSON-first line.
  2. Worker loop: polls Firestore for status="queued", processes chunks or student_text, then writes back status="done" + result.text.

Frontend

• Still calls the backend at POST /upload/assessment and renders evaluation.
• Configure API base via any of:
  • VITE_API_BASE (build time)
  • window.API_BASE (runtime; frontend/public/config.js)
  • localStorage.setItem('API_BASE', 'http(s)://...')

Notes

• Keep Service Account JSON out of git (e.g., store under a private path and point GOOGLE_APPLICATION_CREDENTIALS there).
• Firestore rules should be open only during dev; lock down for production.