#!/usr/bin/env bash
# Benchmark latency on a few representative Hebrew queries.
# Usage: bash scripts/bench_latency.sh

PORT="${PORT:-8000}"
HOST="http://localhost:${PORT}"
# Read API key from env (server seeds it from TAU_RAG_SEED_ADMIN_KEY)
API_KEY="${TAU_RAG_SEED_ADMIN_KEY:-${TAU_RAG_API_KEY:-}}"

# First, sanity-check the server is up
if ! curl -s --max-time 3 "${HOST}/health" > /dev/null 2>&1; then
  echo "✗ Server not responding at ${HOST}"
  echo "  Run 'tail -30 /tmp/tau-rag.log' to see what's happening."
  exit 1
fi
echo "✓ Server alive at ${HOST}"

# Auto-detect which endpoint variant is registered. Public was added in
# the latest patch; admin needs auth.
PUBLIC_PATH="/v1/latency/profile"
ADMIN_PATH="/v1/admin/latency/profile"
ENDPOINT=""
test_resp=$(curl -s --max-time 3 "${HOST}${PUBLIC_PATH}?query=test&n=1" 2>/dev/null)
if [ -n "$test_resp" ] && echo "$test_resp" | grep -q '"stages"'; then
  ENDPOINT="$PUBLIC_PATH"
  echo "  using public endpoint: $ENDPOINT (no auth)"
elif [ -n "$API_KEY" ]; then
  ENDPOINT="$ADMIN_PATH"
  echo "  using admin endpoint: $ENDPOINT (with X-API-Key)"
else
  echo "✗ Neither $PUBLIC_PATH nor $ADMIN_PATH is reachable."
  echo "  Restart the server (it needs the new public route), OR set:"
  echo "    export TAU_RAG_SEED_ADMIN_KEY=<your-key>"
  exit 1
fi
echo

QUERIES=(
  "תום לב במשא ומתן"
  "סעיף 39"
  "אשם תורם"
  "הסכם ממון"
  "פיצויי פיטורים"
)

for q in "${QUERIES[@]}"; do
  echo "=== Query: $q ==="
  encoded=$(python3 -c "import urllib.parse, sys; print(urllib.parse.quote(sys.argv[1]))" "$q")
  if [ "$ENDPOINT" = "$ADMIN_PATH" ] && [ -n "$API_KEY" ]; then
    resp=$(curl -s --max-time 30 -H "X-API-Key: ${API_KEY}" \
            "${HOST}${ENDPOINT}?query=${encoded}&n=3")
  else
    resp=$(curl -s --max-time 30 "${HOST}${ENDPOINT}?query=${encoded}&n=3")
  fi
  echo "$resp" | python3 - <<'PY_END'
import json, sys
raw = sys.stdin.read()
if not raw.strip():
    print("  (empty response — endpoint not responding)")
    sys.exit(0)
try:
    r = json.loads(raw)
except json.JSONDecodeError:
    print("  (non-JSON response)")
    sys.exit(0)
err = r.get("error")
if err:
    print("  ERROR:", err)
    sys.exit(0)
stages = r.get("stages") or {}
items = sorted(stages.items(), key=lambda x: -x[1].get("p50", 0))
for name, s in items:
    p50 = s.get("p50", 0)
    bar = "#" * min(40, int(p50 / 5))
    print("  %-22s p50=%6.1fms  %s" % (name, p50, bar))
bs = r.get("bottleneck_stage")
bp = r.get("bottleneck_p50_ms")
if bs:
    print("  -> bottleneck: %s (%sms)" % (bs, bp))
PY_END
  echo
done