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.

93 lines
3.8 KiB
Python

# backend/app/llm/ollama.py
import json
import httpx
from ..config import get_settings
from .heuristic import HeuristicProvider
from .prompts import CLASSIFY_SYSTEM, build_classify_prompt
from .provider import Classification, LLMProvider
class OllamaProvider(LLMProvider):
name = "ollama"
def __init__(self):
st = get_settings()
self.host = st.ollama_host
self.model = st.ollama_model # 비종속: env 주입값 그대로 사용
self.timeout = st.llm_timeout
def health(self) -> dict:
try:
r = httpx.get(f"{self.host}/api/tags", timeout=3.0)
ok = r.status_code == 200
tags = [m.get("name") for m in r.json().get("models", [])] if ok else []
return {
"reachable": ok, "provider": "ollama", "model": self.model,
"host": self.host, "detail": f"models={tags}",
}
except Exception as e:
return {
"reachable": False, "provider": "ollama", "model": self.model,
"host": self.host, "detail": str(e),
}
def generate_json(self, prompt: str, schema: dict | None = None) -> dict:
# Ollama 구조화 출력: format="json" (모델 비종속). chat API 사용.
# think=False: 추론형 모델의 사고 단계를 끈다(분류는 즉답이 더 적합·빠름).
# 미지원 모델은 무시되거나, 거부 시 except 에서 폴백 처리.
payload = {
"model": self.model,
"messages": [{"role": "user", "content": prompt}],
"stream": False,
"format": "json",
"think": False,
"options": {"temperature": 0.2},
}
r = httpx.post(f"{self.host}/api/chat", json=payload, timeout=self.timeout)
r.raise_for_status()
content = r.json()["message"]["content"]
return json.loads(content)
def classify_capture(self, raw: str, context: dict) -> Classification:
prompt = build_classify_prompt(raw, context)
try:
payload = {
"model": self.model,
"messages": [
{"role": "system", "content": CLASSIFY_SYSTEM},
{"role": "user", "content": prompt},
],
"stream": False,
"format": "json",
"think": False, # 추론 단계 off → 분류 즉답(수초). 미지원 시 except 폴백.
"options": {"temperature": 0.2},
}
r = httpx.post(f"{self.host}/api/chat", json=payload, timeout=self.timeout)
r.raise_for_status()
data = json.loads(r.json()["message"]["content"])
return _coerce(data, raw, context, model=f"ollama:{self.model}")
except Exception:
# 모델 미가용/파싱 실패 → 규칙 폴백 (오프라인에서도 동작)
return HeuristicProvider().classify_capture(raw, context)
def _coerce(data: dict, raw: str, context: dict, model: str) -> Classification:
"""LLM JSON을 Classification으로. 누락 필드는 휴리스틱/기본값 보강."""
h = HeuristicProvider().classify_capture(raw, context) # 폴백 베이스
t = data.get("type") if data.get("type") in ("task", "event", "idea") else h.type
sp = data.get("sphere") if data.get("sphere") in ("work", "life") else h.sphere
return Classification(
type=t, sphere=sp,
project_id=data.get("project_id") or h.project_id,
proj_label=data.get("proj_label") or h.proj_label,
tone=data.get("tone") or h.tone,
due_text=data.get("due_text", h.due_text),
when_text=data.get("when_text", h.when_text),
extra=data.get("extra", h.extra),
reason=data.get("reason") or h.reason,
confidence=float(data.get("confidence", 0.85)),
model=model,
)