1Amogh212/GPT-From-Scratch

1# .env.example — GPT-PRODUCTION-LEVEL backend (app/)2#3# Copy to .env for local development, or set these in your deployment4# platform's environment settings (e.g. Hugging Face Space secrets/variables).5# None of these are read from a .env file automatically by the app itself --6# they're plain os.environ lookups, so export them however your shell/host7# supports (`python-dotenv`, HF Space secrets, docker-compose `environment:`, etc).8 9# Path to the base model checkpoint. If unset, falls back to10# checkpoints/best_model.pt, then checkpoints_tiny/best_model.pt, then an11# untrained dummy model while real weights download in the background.12MODEL_CHECKPOINT=checkpoints/best_model.pt13 14# Name of a LoRA adapter (checkpoints/adapters/<name>.pt) to auto-activate on15# top of the base checkpoint at startup, e.g. an SFT instruction-tuning16# adapter. Optional.17DEFAULT_ADAPTER=sft_v1_small18 19# Comma-separated list of origins allowed by CORS. Must include your deployed20# frontend's exact origin (scheme + host, no path) or browser requests from21# it will be blocked.22ALLOWED_ORIGINS=http://localhost:3000,https://your-frontend.vercel.app23 24# API key for Serper.dev, used for RAG web search (app/search.py). Without25# it, web search silently falls back to scraping DuckDuckGo HTML.26WEB_SEARCH_API_KEY=27 28# Set to "1" only in automated tests -- disables the slowapi rate limiter.29# Do not set this in a real deployment.30TESTING=31 32# Target URL the Streamlit dashboard (app/dashboard.py) talks to.33BACKEND_URL=http://localhost:800034