"""Benchmarks, deren Ergebnisse in Katalog/Registry zurückfließen (Kalibrierung am Deployment-Tag)."""

from __future__ import annotations

import time
from statistics import median
from typing import Any

from .audio import rtf, wav_duration
from .llm import OpenAIChatClient

DEFAULT_TEXT = (
    "Guten Tag. Dies ist ein Test der Sprachausgabe im Cluster. Wir messen, wie lange die Berechnung "
    "im Verhältnis zur Länge des erzeugten Audios dauert. Zahlen wie 1.250,75 Euro und Daten wie der "
    "3. Oktober 2026 sollen korrekt gesprochen werden."
)


def bench_tts(url: str, model: str, *, api_key: str = "", runs: int = 3, text: str | None = None,
              client: OpenAIChatClient | None = None) -> dict[str, Any]:
    client = client or OpenAIChatClient(url, api_key, timeout=600)
    text = text or DEFAULT_TEXT
    results = []
    errors = []
    for _ in range(runs):
        start = time.monotonic()
        try:
            audio = client.speech(model, text)  # type: ignore[attr-defined]
        except Exception as exc:  # noqa: BLE001
            errors.append(str(exc))
            continue
        elapsed = time.monotonic() - start
        duration = wav_duration(audio)
        results.append({"seconds": round(elapsed, 2), "audio_seconds": round(duration, 2),
                        "rtf": round(rtf(elapsed, duration), 3) if duration else None})
    rtfs = [r["rtf"] for r in results if r["rtf"] is not None]
    return {"model": model, "chars": len(text), "runs": results, "errors": errors,
            "rtf_median": round(median(rtfs), 3) if rtfs else None}