fly.toml

# Fly.io config for the RAG service (the Next.js app deploys to Vercel).
#
# Deploy:
#   fly launch --no-deploy    # claims the app name / picks a region, keeps this file
#   fly deploy
#
# The app name below must be globally unique on Fly — `fly launch` will offer to
# change it if it is taken.

app = "rag-engine-demo"
primary_region = "iad"   # change to whichever region is nearest you

[build]
  dockerfile = "Dockerfile"

[env]
  # A fresh machine has nobody to run scripts/seed.py, so the service indexes
  # its bundled sample corpus on startup. Also set in the Dockerfile; kept here
  # so the behavior is visible in the deploy config.
  RAG_SEED_ON_STARTUP = "1"
  RAG_UPLOADS_ENABLED = "0"
  RAG_CORS_ORIGINS = ""
  RAG_CLAUDE_TIMEOUT_SECONDS = "30"
  RAG_MAX_TITLE_CHARS = "200"
  RAG_MAX_DOCUMENT_CHARS = "1000000"
  RAG_MAX_QUESTION_CHARS = "2000"
  RAG_MAX_REQUEST_BYTES = "4194304"

[http_service]
  internal_port = 8080
  force_https = true

  # Scale to zero when idle. "suspend" snapshots the machine's memory rather
  # than killing it, so the in-memory HNSW index survives a resume and wake is
  # ~1-2s. Startup seeding covers the case where the machine fully stops.
  # If your flyctl is too old to accept "suspend", use "stop" — the index is
  # then rebuilt from the sample corpus on each cold start, which is fine.
  auto_stop_machines = "suspend"
  auto_start_machines = true
  min_machines_running = 0

  [[http_service.checks]]
    grace_period = "10s"
    interval = "30s"
    timeout = "5s"
    method = "GET"
    path = "/healthz"

[[vm]]
  # 512MB is ample: the image has no torch, because the service runs the
  # dependency-free hashed embedder rather than sentence-transformers.
  memory = "512mb"
  cpu_kind = "shared"
  cpus = 1