Что логировать в продакшне, чтобы разбирать сбои языковой модели

from dataclasses import dataclass, field, asdict
from datetime import datetime, timezone
import uuid
import time
import json
from openai import OpenAI

client = OpenAI()

@dataclass
class LlmLog:
    request_id: str
    timestamp: str
    prompt_version: str
    model: str
    temperature: float
    user_text: str
    answer: str
    prompt_tokens: int
    completion_tokens: int
    elapsed_ms: int
    error: str = ""

def ask_with_log(user_text: str, prompt_version: str = "1.0.0") -> str:
    request_id = str(uuid.uuid4())
    start = time.monotonic()
    error = ""
    answer = ""
    usage = None

    try:
        resp = client.chat.completions.create(
            model="gpt-4.1",
            temperature=0.3,
            max_tokens=512,
            messages=[
                {"role": "system", "content": "Ты помощник службы поддержки."},
                {"role": "user", "content": user_text}
            ]
        )
        answer = resp.choices[0].message.content
        usage = resp.usage
    except Exception as e:
        error = str(e)
    finally:
        elapsed = int((time.monotonic() - start) * 1000)

    log = LlmLog(
        request_id=request_id,
        timestamp=datetime.now(timezone.utc).isoformat(),
        prompt_version=prompt_version,
        model="gpt-4.1",
        temperature=0.3,
        user_text=user_text,
        answer=answer,
        prompt_tokens=usage.prompt_tokens if usage else 0,
        completion_tokens=usage.completion_tokens if usage else 0,
        elapsed_ms=elapsed,
        error=error
    )

    # В продакшне — пишите в хранилище (Postgres, ClickHouse, S3)
    print(json.dumps(asdict(log), ensure_ascii=False))

    if error:
        raise RuntimeError(error)

    return answer