Bloomberg Terminal for Burgundy — pricing intelligence for all 34 Grand Cru climats.

Live: burgreport.com API: burgreport-production.up.railway.app (/docs)

• Backend: Python 3.12 + FastAPI on Railway. Routers for search, wines, vintages.
• AI layer: OpenAI Responses API with the web_search tool, falling back to Tavily search over public merchant listings. Prices are unvalidated estimates parsed from public web pages — not a licensed feed and not a Wine-Searcher integration.
• Cache: Supabase Postgres, 24-hour TTL on price snapshots.
• Content: Airtable REST API for curated climat descriptions, producers, and pairings (1-hour in-memory cache).
• Frontend: Lovable-managed React SPA on Cloudflare. Not tracked in this repo.
• Refresh: Nightly Railway cron (0 3 * * *) iterates 34 climats × 5 vintages = 170 price lookups per night.

backend/                  Python FastAPI app — canonical code path
  README.md               Backend-specific setup
  Procfile                Railway entry point
  routers/                FastAPI routes
  services/               OpenAI, Supabase, Airtable, reference data
  models/                 Pydantic models + truth-status enum
  refresh_prices.py       Nightly cron entry point
  tests/                  Unit tests
supabase/migrations/      Tracked DB migrations
docs/                     DATA_STATUS_MODEL, SUPABASE_SCHEMA, PRODUCTION_READINESS_CHECKLIST
burgreport-next-vercel-v2/  LEGACY — see CLAIM.md inside; not part of canonical stack

See backend/README.md for backend setup, environment variables, and local validation steps.

Required env vars on Railway: OPENAI_API_KEY, SUPABASE_URL, SUPABASE_KEY, AIRTABLE_TOKEN, AIRTABLE_BASE_ID. OPENAI_SEARCH_MODEL defaults to gpt-4.1-mini for price lookups using web search. OPENAI_MODEL defaults to gpt-4.1-mini for non-search fallback context; set either to gpt-5 or an o-series model to enable reasoning effort.

Live with caveats — see docs/PRODUCTION_READINESS_CHECKLIST.md for the current blockers. Recent audit + remediation findings are tracked in private maintainer notes.

The backend never fabricates prices, merchants, source counts, confidence scores, or comparables. Missing values stay unavailable rather than being substituted. See docs/DATA_STATUS_MODEL.md.