Airfare Intelligence MVP

Hackathon MVP: route-aware sitemap generation, permitted Crawl4AI extraction hooks, SHA-256 deduplication, Ollama/Llama 3.1 readiness, FastAPI, and a React + Plotly dashboard.

Local

uv sync
uv run uvicorn app:app --app-dir src --reload
cd web && npm install && npm run dev

Open http://localhost:5173. Set VITE_API_URL=http://localhost:8000 if the API is not local. Optional crawler/model setup: uv sync --extra crawler --extra local-llm, install Crawl4AI browser dependencies, install Ollama, then ollama pull llama3.1:latest.

Railway deployment

1. Create a Railway project from this repository.
2. Create an API service with root /, Dockerfile Dockerfile, and variables PERMITTED_DOMAINS, CORS_ORIGINS, OLLAMA_BASE_URL, and OLLAMA_MODEL=llama3.1:latest. Railway provides PORT; do not hardcode it. Healthcheck: /health.
3. Deploy. The API start command is already in the Dockerfile. Copy its public URL.
4. Create a second service for the React app, root directory web, Dockerfile web/Dockerfile, and VITE_API_URL=<API public URL>. Expose port 4173 and use / as the healthcheck.
5. Redeploy the web service after changing VITE_API_URL because Vite embeds it at build time.

Before deployment: replace demo data with a database/object store for persistence, set an explicit permitted-domain allowlist, set CORS_ORIGINS to the web URL instead of *, install Crawl4AI browser dependencies in the API image, and add rate limits/timeouts. A laptop Ollama process is not reachable from Railway; deploy Ollama on a reachable GPU/VM service and set its HTTPS URL, or leave it unavailable and use the visible fallback warning. Never crawl login pages, CAPTCHAs, disallowed robots, or restricted domains.

API flow

POST /api/v1/sitemap accepts domain, origin, and destination, then returns route query URLs. POST /api/v1/scrape queues those URLs and returns normalized records; replace the marked adapter with AsyncWebCrawler, CrawlerRunConfig, DefaultMarkdownGenerator, and JSON extraction on permitted domains. Store content_hash=sha256(canonical_url + cleaned_content) to prevent duplicate fare records. The UI exposes booking URLs, route history, airline comparison, and crawl/model status.

Streamlit remains available as a fallback: uv run streamlit run src/dashboard.py.