Capstone: ship the RAG app

#!/usr/bin/env bash
# Start NEW_IMAGE next to the running version, wait for /ready, then switch over.
# If it never becomes ready, remove it: the old version keeps serving.
set -euo pipefail
NEW_IMAGE="$1"; shift
 
echo "==> starting candidate $NEW_IMAGE on port 8001"
docker run -d --name rag-candidate -p 8001:8000 "$@" "$NEW_IMAGE" > /dev/null
 
for i in $(seq 1 15); do
  if curl -fs http://localhost:8001/ready > /dev/null; then
    echo "==> candidate ready after ${i}s: $(curl -s http://localhost:8001/ready)"
    docker rm -f rag-live > /dev/null 2>&1 || true
    docker rm -f rag-candidate > /dev/null
    docker run -d --name rag-live -p 8000:8000 "$@" "$NEW_IMAGE" > /dev/null
    echo "==> promoted: live is now $NEW_IMAGE"
    exit 0
  fi
  sleep 1
done
 
echo "==> candidate NOT ready after 15s: $(curl -s http://localhost:8001/ready || true)"
docker rm -f rag-candidate > /dev/null
echo "==> ROLLED BACK: live version unchanged: $(curl -s http://localhost:8000/health || echo 'none running')"
exit 1

# Deploy with health check
==> starting candidate ghcr.io/<owner>/all-about-ai/rag-app:31fab5b on port 8001
==> candidate ready after 3s: {"ready":true,"checks":{"index_loaded":true,"provider_configured":true},"version":"31fab5b"}
==> promoted: live is now ghcr.io/<owner>/all-about-ai/rag-app:31fab5b
 
# Rollback drill (same image, started with LLM_PROVIDER=anthropic and no key)
==> starting candidate ghcr.io/<owner>/all-about-ai/rag-app:31fab5b on port 8001
==> candidate NOT ready after 15s: {"ready":false,"checks":{"index_loaded":true,"provider_configured":false},"version":"31fab5b"}
==> ROLLED BACK: live version unchanged: {"status":"ok","chunks":12,"version":"31fab5b"}
drill passed: live version still serving -> {"status":"ok","chunks":12,"version":"31fab5b"}