You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

109 lines
3.9 KiB
Python

# backend/app/llm/openai_compat.py
# OpenAI 호환 Chat Completions 엔드포인트용 프로바이더.
# OpenAI 본가뿐 아니라 OpenRouter·Groq·Together·vLLM·LM Studio·Gemini(OpenAI 호환 베이스) 등
# `/v1/chat/completions` + `/v1/models` 규약을 따르는 모든 엔드포인트를 host(base_url)로 받는다.
import httpx
from ..runtime_config import effective_llm
from .ollama import _coerce, _lenient_json
from .prompts import CLASSIFY_SYSTEM, build_classify_prompt
from .provider import Classification, LLMProvider
class OpenAICompatProvider(LLMProvider):
name = "openai"
def __init__(
self,
host: str | None = None,
model: str | None = None,
api_key: str | None = None,
timeout: float | None = None,
):
eff = effective_llm()
# host = base_url(예: https://api.openai.com/v1). 끝 슬래시 정규화.
self.base = (host or eff.host).rstrip("/")
self.model = model or eff.model
self.api_key = api_key if api_key is not None else eff.api_key
self.timeout = timeout if timeout is not None else eff.timeout
def _headers(self) -> dict:
h = {"Content-Type": "application/json"}
if self.api_key:
h["Authorization"] = f"Bearer {self.api_key}"
return h
def _chat(self, messages: list[dict]) -> str:
body = {
"model": self.model,
"messages": messages,
"temperature": 0.2,
"response_format": {"type": "json_object"},
}
try:
r = httpx.post(
f"{self.base}/chat/completions",
json=body,
headers=self._headers(),
timeout=self.timeout,
)
r.raise_for_status()
except httpx.HTTPStatusError as e:
# 일부 호환 서버는 response_format 을 모름(400) → 빼고 1회 재시도.
if e.response is not None and e.response.status_code == 400:
body.pop("response_format", None)
r = httpx.post(
f"{self.base}/chat/completions",
json=body,
headers=self._headers(),
timeout=self.timeout,
)
r.raise_for_status()
else:
raise
return r.json()["choices"][0]["message"]["content"]
def list_models(self) -> list[str]:
try:
r = httpx.get(f"{self.base}/models", headers=self._headers(), timeout=5.0)
if r.status_code != 200:
return []
data = r.json().get("data", [])
return [m.get("id", "") for m in data if m.get("id")]
except Exception:
return []
def health(self) -> dict:
try:
r = httpx.get(f"{self.base}/models", headers=self._headers(), timeout=5.0)
ok = r.status_code == 200
return {
"reachable": ok,
"provider": self.name,
"model": self.model,
"host": self.base,
"detail": "ok" if ok else f"HTTP {r.status_code}",
}
except Exception as e:
return {
"reachable": False,
"provider": self.name,
"model": self.model,
"host": self.base,
"detail": str(e),
}
def generate_json(self, prompt: str, schema: dict | None = None) -> dict:
content = self._chat([{"role": "user", "content": prompt}])
return _lenient_json(content)
def classify_capture(self, raw: str, context: dict) -> Classification:
prompt = build_classify_prompt(raw, context)
content = self._chat(
[
{"role": "system", "content": CLASSIFY_SYSTEM},
{"role": "user", "content": prompt},
]
)
return _coerce(_lenient_json(content), raw, context, model=f"openai:{self.model}")