279 lines
10 KiB
Python
279 lines
10 KiB
Python
from __future__ import annotations
|
|
|
|
from typing import Any, Dict, List, Optional, Tuple
|
|
import time
|
|
import logging
|
|
|
|
from openai import OpenAI, AsyncOpenAI
|
|
from tenacity import retry, stop_after_attempt, wait_exponential, retry_if_exception_type
|
|
|
|
from .config import extractor_llm_config, settings
|
|
|
|
|
|
class LLMClient:
|
|
"""Cliente unificado para llamadas Chat Completions (SDK OpenAI).
|
|
|
|
- Centraliza api_key, base_url, modelo y timeout desde configuracion.
|
|
- Expone helpers para respuestas de texto y JSON.
|
|
- Registra métricas mínimas en `domain.api_calls` (start/end/error + usage).
|
|
- Aplica reintentos ante timeouts.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
api_key: Optional[str] = None,
|
|
base_url: Optional[str] = None,
|
|
default_model: Optional[str] = None,
|
|
request_timeout_s: Optional[float] = None,
|
|
) -> None:
|
|
default_config = extractor_llm_config()
|
|
self.api_key = api_key if api_key is not None else str(default_config["api_key"])
|
|
self.base_url = base_url if base_url is not None else str(default_config["base_url"])
|
|
self.default_model = default_model if default_model is not None else str(default_config["model"])
|
|
self.request_timeout_s = request_timeout_s or float(default_config["timeout_s"])
|
|
self._client = self._build_sync_client()
|
|
self._async_client = self._build_async_client()
|
|
self._api_logger = logging.getLogger("domain.api_calls")
|
|
|
|
def _build_sync_client(self) -> OpenAI:
|
|
kwargs: Dict[str, Any] = {"api_key": self.api_key}
|
|
if self.base_url:
|
|
kwargs["base_url"] = self.base_url
|
|
return OpenAI(**kwargs)
|
|
|
|
def _build_async_client(self) -> AsyncOpenAI:
|
|
kwargs: Dict[str, Any] = {"api_key": self.api_key}
|
|
if self.base_url:
|
|
kwargs["base_url"] = self.base_url
|
|
return AsyncOpenAI(**kwargs)
|
|
|
|
def update_config(
|
|
self,
|
|
*,
|
|
api_key: Optional[str] = None,
|
|
base_url: Optional[str] = None,
|
|
default_model: Optional[str] = None,
|
|
request_timeout_s: Optional[float] = None,
|
|
) -> None:
|
|
if api_key is not None:
|
|
self.api_key = api_key
|
|
if base_url is not None:
|
|
self.base_url = base_url
|
|
if default_model is not None:
|
|
self.default_model = default_model
|
|
if request_timeout_s is not None:
|
|
self.request_timeout_s = request_timeout_s
|
|
self._client = self._build_sync_client()
|
|
self._async_client = self._build_async_client()
|
|
|
|
def list_models(self) -> List[str]:
|
|
"""Lista modelos publicados por el backend compatible con OpenAI."""
|
|
models = self._client.models.list()
|
|
return sorted(
|
|
str(getattr(model, "id", ""))
|
|
for model in getattr(models, "data", [])
|
|
if getattr(model, "id", None)
|
|
)
|
|
|
|
@retry(
|
|
stop=stop_after_attempt(settings.retry.max_attempts),
|
|
wait=wait_exponential(multiplier=settings.retry.wait_min_s, max=settings.retry.wait_max_s),
|
|
retry=retry_if_exception_type((TimeoutError,)),
|
|
reraise=True,
|
|
)
|
|
def _create_completion(self, *, messages: List[Dict[str, Any]], model: str, timeout_s: float, **extra) -> Any:
|
|
return self._client.chat.completions.create(
|
|
messages=messages,
|
|
model=model,
|
|
timeout=timeout_s,
|
|
**extra,
|
|
)
|
|
|
|
@retry(
|
|
stop=stop_after_attempt(settings.retry.max_attempts),
|
|
wait=wait_exponential(multiplier=settings.retry.wait_min_s, max=settings.retry.wait_max_s),
|
|
retry=retry_if_exception_type((TimeoutError,)),
|
|
reraise=True,
|
|
)
|
|
async def _create_completion_async(self, *, messages: List[Dict[str, Any]], model: str, timeout_s: float, **extra) -> Any:
|
|
return await self._async_client.chat.completions.create(
|
|
messages=messages,
|
|
model=model,
|
|
timeout=timeout_s,
|
|
**extra,
|
|
)
|
|
|
|
def _log_start(self, *, component: str, model: str, timeout_s: float, messages_count: int) -> None:
|
|
try:
|
|
self._api_logger.info(
|
|
"api_call_start",
|
|
extra={
|
|
"component": component,
|
|
"base_url": self.base_url,
|
|
"model": model,
|
|
"timeout_s": timeout_s,
|
|
"messages": messages_count,
|
|
},
|
|
)
|
|
except Exception:
|
|
pass
|
|
|
|
def _log_end(self, *, component: str, model: str, timeout_s: float, elapsed_ms: int, usage: Optional[Dict[str, Any]]) -> None:
|
|
try:
|
|
self._api_logger.info(
|
|
"api_call_end",
|
|
extra={
|
|
"component": component,
|
|
"base_url": self.base_url,
|
|
"model": model,
|
|
"timeout_s": timeout_s,
|
|
"elapsed_ms": elapsed_ms,
|
|
"prompt_tokens": (usage or {}).get("prompt_tokens"),
|
|
"completion_tokens": (usage or {}).get("completion_tokens"),
|
|
"total_tokens": (usage or {}).get("total_tokens"),
|
|
},
|
|
)
|
|
except Exception:
|
|
pass
|
|
|
|
def _log_error(self, *, component: str, model: str, timeout_s: float, elapsed_ms: Optional[int], error: Exception) -> None:
|
|
try:
|
|
self._api_logger.info(
|
|
"api_call_error",
|
|
extra={
|
|
"component": component,
|
|
"base_url": self.base_url,
|
|
"model": model,
|
|
"timeout_s": timeout_s,
|
|
"elapsed_ms": (elapsed_ms if elapsed_ms is not None else None),
|
|
"error": str(error),
|
|
},
|
|
)
|
|
except Exception:
|
|
pass
|
|
|
|
@staticmethod
|
|
def _extract_usage(obj: Any) -> Optional[Dict[str, Any]]:
|
|
try:
|
|
usage = getattr(obj, "usage", None)
|
|
if usage is None:
|
|
return None
|
|
usage_dict: Dict[str, Any] = {
|
|
"prompt_tokens": getattr(usage, "prompt_tokens", None),
|
|
"completion_tokens": getattr(usage, "completion_tokens", None),
|
|
"total_tokens": getattr(usage, "total_tokens", None),
|
|
}
|
|
details = getattr(usage, "prompt_tokens_details", None)
|
|
if details is not None:
|
|
try:
|
|
usage_dict["prompt_tokens_details"] = dict(details) # type: ignore[arg-type]
|
|
except Exception:
|
|
try:
|
|
usage_dict["prompt_tokens_details"] = {
|
|
"cached_tokens": getattr(details, "cached_tokens", None),
|
|
"audio_tokens": getattr(details, "audio_tokens", None),
|
|
}
|
|
except Exception:
|
|
pass
|
|
return usage_dict
|
|
except Exception:
|
|
return None
|
|
|
|
def chat_text(
|
|
self,
|
|
messages: List[Dict[str, Any]],
|
|
*,
|
|
model: Optional[str] = None,
|
|
timeout_s: Optional[float] = None,
|
|
component: str = "llm_client",
|
|
) -> Tuple[str, Optional[Dict[str, Any]]]:
|
|
"""Devuelve (texto, usage)."""
|
|
eff_model = model or self.default_model
|
|
eff_timeout = float(timeout_s if timeout_s is not None else self.request_timeout_s)
|
|
self._log_start(component=component, model=eff_model, timeout_s=eff_timeout, messages_count=len(messages or []))
|
|
start = time.time()
|
|
try:
|
|
chat_completion = self._create_completion(
|
|
messages=messages,
|
|
model=eff_model,
|
|
timeout_s=eff_timeout,
|
|
)
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
usage = self._extract_usage(chat_completion)
|
|
self._log_end(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, usage=usage)
|
|
text: str = chat_completion.choices[0].message.content # type: ignore[assignment]
|
|
return text, usage
|
|
except Exception as ex:
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
self._log_error(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, error=ex)
|
|
raise
|
|
async def chat_text_async(
|
|
self,
|
|
messages: List[Dict[str, Any]],
|
|
*,
|
|
model: Optional[str] = None,
|
|
timeout_s: Optional[float] = None,
|
|
component: str = "llm_client",
|
|
) -> Tuple[str, Optional[Dict[str, Any]]]:
|
|
"""Devuelve (texto, usage) de forma asíncrona."""
|
|
eff_model = model or self.default_model
|
|
eff_timeout = float(timeout_s if timeout_s is not None else self.request_timeout_s)
|
|
self._log_start(component=component, model=eff_model, timeout_s=eff_timeout, messages_count=len(messages or []))
|
|
start = time.time()
|
|
try:
|
|
chat_completion = await self._create_completion_async(
|
|
messages=messages,
|
|
model=eff_model,
|
|
timeout_s=eff_timeout,
|
|
)
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
usage = self._extract_usage(chat_completion)
|
|
self._log_end(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, usage=usage)
|
|
text: str = chat_completion.choices[0].message.content # type: ignore[assignment]
|
|
return text, usage
|
|
except Exception as ex:
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
self._log_error(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, error=ex)
|
|
raise
|
|
def chat_json(
|
|
self,
|
|
messages: List[Dict[str, Any]],
|
|
*,
|
|
model: Optional[str] = None,
|
|
response_format: str = "json_object",
|
|
timeout_s: Optional[float] = None,
|
|
component: str = "llm_client",
|
|
max_completion_tokens: Optional[int] = None,
|
|
) -> Tuple[Dict[str, Any], Optional[Dict[str, Any]]]:
|
|
"""Devuelve (objeto JSON parseado, usage)."""
|
|
import json
|
|
eff_model = model or self.default_model
|
|
eff_timeout = float(timeout_s if timeout_s is not None else self.request_timeout_s)
|
|
self._log_start(component=component, model=eff_model, timeout_s=eff_timeout, messages_count=len(messages or []))
|
|
start = time.time()
|
|
try:
|
|
completion_kwargs: Dict[str, Any] = {
|
|
"response_format": {"type": response_format},
|
|
}
|
|
if max_completion_tokens is not None:
|
|
completion_kwargs["max_completion_tokens"] = max_completion_tokens
|
|
chat_completion = self._create_completion(
|
|
messages=messages,
|
|
model=eff_model,
|
|
timeout_s=eff_timeout,
|
|
**completion_kwargs,
|
|
)
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
usage = self._extract_usage(chat_completion)
|
|
self._log_end(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, usage=usage)
|
|
content = chat_completion.choices[0].message.content # type: ignore[assignment]
|
|
obj = json.loads(content)
|
|
return obj, usage
|
|
except Exception as ex:
|
|
elapsed_ms = int((time.time() - start) * 1000)
|
|
self._log_error(component=component, model=eff_model, timeout_s=eff_timeout, elapsed_ms=elapsed_ms, error=ex)
|
|
raise
|
|
|
|
|
|
__all__ = ["LLMClient"]
|