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.
97 lines
3.6 KiB
Python
97 lines
3.6 KiB
Python
# backend/app/llm/gemini.py
|
|
# Google Gemini (Generative Language API) 프로바이더.
|
|
# 엔드포인트: {base}/models/{model}:generateContent?key=API_KEY (base 기본 v1beta).
|
|
# host 오버레이가 ollama 용 값으로 남아 있을 수 있어, gemini 답지 않은 host 는 기본 base 로 무시.
|
|
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
|
|
|
|
DEFAULT_BASE = "https://generativelanguage.googleapis.com/v1beta"
|
|
|
|
|
|
def _resolve_base(host: str | None) -> str:
|
|
h = (host or "").rstrip("/")
|
|
# 비어 있거나 ollama 류 host(11434/ollama/localhost)면 gemini 기본 base 사용.
|
|
if not h or ":11434" in h or "ollama" in h or "localhost" in h or "127.0.0.1" in h:
|
|
return DEFAULT_BASE
|
|
return h
|
|
|
|
|
|
class GeminiProvider(LLMProvider):
|
|
name = "gemini"
|
|
|
|
def __init__(
|
|
self,
|
|
host: str | None = None,
|
|
model: str | None = None,
|
|
api_key: str | None = None,
|
|
timeout: float | None = None,
|
|
):
|
|
eff = effective_llm()
|
|
self.base = _resolve_base(host if host is not None else eff.host)
|
|
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 _generate(self, text: str) -> str:
|
|
body = {
|
|
"contents": [{"role": "user", "parts": [{"text": text}]}],
|
|
"generationConfig": {"temperature": 0.2, "responseMimeType": "application/json"},
|
|
}
|
|
r = httpx.post(
|
|
f"{self.base}/models/{self.model}:generateContent",
|
|
params={"key": self.api_key},
|
|
json=body,
|
|
timeout=self.timeout,
|
|
)
|
|
r.raise_for_status()
|
|
cands = r.json().get("candidates", [])
|
|
if not cands:
|
|
return "{}"
|
|
parts = cands[0].get("content", {}).get("parts", [])
|
|
return "".join(p.get("text", "") for p in parts)
|
|
|
|
def list_models(self) -> list[str]:
|
|
try:
|
|
r = httpx.get(f"{self.base}/models", params={"key": self.api_key}, timeout=5.0)
|
|
if r.status_code != 200:
|
|
return []
|
|
out = []
|
|
for m in r.json().get("models", []):
|
|
if "generateContent" in (m.get("supportedGenerationMethods") or []):
|
|
out.append((m.get("name") or "").removeprefix("models/"))
|
|
return [m for m in out if m]
|
|
except Exception:
|
|
return []
|
|
|
|
def health(self) -> dict:
|
|
try:
|
|
r = httpx.get(f"{self.base}/models", params={"key": self.api_key}, 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:
|
|
return _lenient_json(self._generate(prompt))
|
|
|
|
def classify_capture(self, raw: str, context: dict) -> Classification:
|
|
prompt = f"{CLASSIFY_SYSTEM}\n\n{build_classify_prompt(raw, context)}"
|
|
data = _lenient_json(self._generate(prompt))
|
|
return _coerce(data, raw, context, model=f"gemini:{self.model}")
|