From 867ff7993f6be451dea2bb8a2649aba0e28d6ea2 Mon Sep 17 00:00:00 2001 From: Junseo1026 Date: Sun, 21 Sep 2025 16:03:16 +0900 Subject: [PATCH] summary --- Makefile | 2 +- models/checklist.py | 26 ++ requirements.txt | 1 + routers/checklist.py | 47 ++++ routers/note.py | 46 +++- schemas/checklist.py | 8 + utils/llm.py | 598 +++++++++++++++++++++++++++++++++++++++++-- 7 files changed, 697 insertions(+), 31 deletions(-) create mode 100644 models/checklist.py create mode 100644 routers/checklist.py create mode 100644 schemas/checklist.py diff --git a/Makefile b/Makefile index 3840c38..dce5c46 100644 --- a/Makefile +++ b/Makefile @@ -1,3 +1,3 @@ # Makefile run: - uvicorn main:app --host 0.0.0.0 --port 8080 --reload --http h11 + uvicorn main:app --host 0.0.0.0 --port 8080 --reload --http h11 \ No newline at end of file diff --git a/models/checklist.py b/models/checklist.py new file mode 100644 index 0000000..e7e2399 --- /dev/null +++ b/models/checklist.py @@ -0,0 +1,26 @@ +from sqlalchemy import Column, Integer, String, Enum, TIMESTAMP, text +from sqlalchemy.orm import relationship + +class User(Base): + __tablename__ = "user" + + u_id = Column(Integer, primary_key=True, autoincrement=True) # PK + id = Column(String(50), unique=True, nullable=False) # 로그인 ID + email = Column(String(150), unique=True, nullable=False) + password = Column(String(255), nullable=False) + provider = Column( + Enum("local", "google", "kakao", "naver", name="provider_enum"), + nullable=False, + server_default="local", + ) + created_at = Column( + TIMESTAMP, nullable=False, server_default=text("CURRENT_TIMESTAMP") + ) + updated_at = Column( + TIMESTAMP, + nullable=False, + server_default=text("CURRENT_TIMESTAMP ON UPDATE CURRENT_TIMESTAMP"), + ) + + # 역참조: checklist 목록 + checklists = relationship("Checklist", back_populates="user", cascade="all, delete-orphan") \ No newline at end of file diff --git a/requirements.txt b/requirements.txt index 495ed74..27e7ec0 100644 --- a/requirements.txt +++ b/requirements.txt @@ -31,3 +31,4 @@ langchain-community langchain-core langchain-openai langchain-ollama +bitsandbytes \ No newline at end of file diff --git a/routers/checklist.py b/routers/checklist.py new file mode 100644 index 0000000..26a41b1 --- /dev/null +++ b/routers/checklist.py @@ -0,0 +1,47 @@ +# 변경/설명: +# - POST /checklists : 생성 전용 +# - PATCH /checklists/{id}/clear : is_clear 0/1 설정 +# - get_current_user는 user.u_id 제공 가정 +from fastapi import APIRouter, Depends, HTTPException, status +from sqlalchemy.orm import Session + +from app.schemas.checklist import ChecklistCreate, ChecklistClearUpdate +from app.models import Checklist # Checklist(u_id, checklist_title, is_clear) +from app.dependencies import get_db, get_current_user + +router = APIRouter(prefix="/checklists", tags=["checklists"]) + +@router.post("", status_code=status.HTTP_201_CREATED) +def create_checklist( + req: ChecklistCreate, + db: Session = Depends(get_db), + user=Depends(get_current_user), +): + obj = Checklist( + u_id=user.u_id, # ← 프로젝트의 사용자 키에 맞게 + checklist_title=req.checklist_title, + is_clear=0 # 기본 0(미완) + ) + db.add(obj) + db.commit() + db.refresh(obj) + return {"id": obj.id, "checklist_title": obj.checklist_title, "is_clear": obj.is_clear} + +@router.patch("/{checklist_id}/clear") +def set_clear_state( + checklist_id: int, + req: ChecklistClearUpdate, # {"is_clear": 0 | 1} + db: Session = Depends(get_db), + user=Depends(get_current_user), +): + obj = ( + db.query(Checklist) + .filter(Checklist.id == checklist_id, Checklist.u_id == user.u_id) + .first() + ) + if not obj: + raise HTTPException(status_code=404, detail="Checklist not found") + obj.is_clear = int(req.is_clear) # 0/1 저장 + db.commit() + db.refresh(obj) + return {"id": obj.id, "is_clear": obj.is_clear} \ No newline at end of file diff --git a/routers/note.py b/routers/note.py index 552a61a..3e7f1cc 100644 --- a/routers/note.py +++ b/routers/note.py @@ -1,6 +1,6 @@ import os from dotenv import load_dotenv -from fastapi import APIRouter, Depends, HTTPException, BackgroundTasks +from fastapi import APIRouter, Depends, HTTPException, BackgroundTasks, Query from fastapi.responses import StreamingResponse from sqlalchemy.orm import Session from typing import List @@ -145,13 +145,24 @@ def toggle_favorite( return note def save_summary(note_id: int, text: str): + """Deprecated: kept for reference. No longer overwrites original note.""" db2 = SessionLocal() try: tgt = db2.query(Note).filter(Note.id == note_id).first() - if tgt: - tgt.content = text - tgt.updated_at = datetime.utcnow() - db2.commit() + if not tgt: + return + # Create a new note in the same folder with title '요약' + title = (tgt.title or "").strip() + "요약" + if len(title) > 255: + title = title[:255] + new_note = Note( + user_id=tgt.user_id, + folder_id=tgt.folder_id, + title=title, + content=text, + ) + db2.add(new_note) + db2.commit() finally: db2.close() @@ -159,6 +170,8 @@ def save_summary(note_id: int, text: str): async def summarize_stream_langchain( note_id: int, background_tasks: BackgroundTasks, + domain: str | None = Query(default=None, description="meeting | code | paper | general | auto(None)"), + longdoc: bool = Query(default=True, description="Enable long-document map→reduce"), db: Session = Depends(get_db), user = Depends(get_current_user) ): @@ -168,15 +181,32 @@ async def summarize_stream_langchain( async def event_gen(): parts = [] - async for sse in stream_summary_with_langchain(note.content): + async for sse in stream_summary_with_langchain(note.content, domain=domain, longdoc=longdoc): parts.append(sse.removeprefix("data: ").strip()) yield sse.encode() full = "".join(parts).strip() if full: - background_tasks.add_task(save_summary, note.id, full) + # Create a new summary note in the same folder with title '요약' + title = (note.title or "").strip() + "요약" + if len(title) > 255: + title = title[:255] + new_note = Note( + user_id=user.u_id, + folder_id=note.folder_id, + title=title, + content=full, + ) + db.add(new_note) + db.commit() + db.refresh(new_note) + try: + # Optional: notify created note id + yield f"data: SUMMARY_NOTE_ID:{new_note.id}\n\n".encode() + except Exception: + pass return StreamingResponse( event_gen(), media_type="text/event-stream", headers={"Cache-Control": "no-cache"} - ) \ No newline at end of file + ) diff --git a/schemas/checklist.py b/schemas/checklist.py new file mode 100644 index 0000000..04aa26f --- /dev/null +++ b/schemas/checklist.py @@ -0,0 +1,8 @@ + +from pydantic import BaseModel, conint + +class ChecklistCreate(BaseModel): + checklist_title: str + +class ChecklistClearUpdate(BaseModel): + is_clear: conint(ge=0, le=1) # 0 또는 1만 허용 \ No newline at end of file diff --git a/utils/llm.py b/utils/llm.py index 31946c4..fb39e40 100644 --- a/utils/llm.py +++ b/utils/llm.py @@ -1,50 +1,604 @@ +# [CHANGED] 마크다운 보정 + Markdown 섹션 포맷 + 선택적 웹 보강(위키) + 동일 언어 요약을 포함한 전체 코드 from langchain.callbacks import AsyncIteratorCallbackHandler from langchain_ollama import ChatOllama from langchain.schema import HumanMessage, SystemMessage -import re, asyncio +import re, asyncio, os, threading, json, time +import torch +from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer +# [CHANGED] 웹 보강용 (옵션) +import requests +from urllib.parse import quote as _urlquote + +# =============== 필터: 사고과정 유사 문장 =============== _THOUGHT_PAT = re.compile( r"^\s*(okay|let\s*me|i\s*need\s*to|first[, ]|then[, ]|next[, ]|in summary|먼저|그\s*다음|요약하면)", re.I, ) -async def stream_summary_with_langchain(text: str): +# =============== 마크다운 보정 유틸 =============== +# [CHANGED] 스트리밍 중 붙어버린 헤더/불릿을 자동 교정 +_MD_CODE_FENCE = re.compile(r"(```.*?```|`[^`]*`)", re.S) + +def _format_md_stream(s: str) -> str: """ - LangChain + Ollama에서 토큰을 비동기로 받아 - SSE("data: ...\\n\\n") 형식으로 yield 하는 async generator + 스트리밍 중 합쳐진 마크다운을 사람 읽기 좋게 보정. + - 헤더(#~######) 앞뒤 빈 줄 보장 + - 헤더 해시 뒤 공백 보장 + - 불릿(- )/번호 목록은 줄 시작으로 강제 + - 코드펜스/인라인코드는 건드리지 않음 """ + if not s: + return s + + parts = [] + last = 0 + for m in _MD_CODE_FENCE.finditer(s): + chunk = s[last:m.start()] + parts.append(_format_md_plain(chunk)) + parts.append(m.group(0)) # 코드펜스 원형 유지 + last = m.end() + parts.append(_format_md_plain(s[last:])) + + out = "".join(parts) + out = re.sub(r"\n{3,}", "\n\n", out) # 과도한 빈 줄 축소 + return out + +def _format_md_plain(s: str) -> str: + # 헤더 해시 뒤 공백 보장: "##개요" -> "## 개요" + s = re.sub(r"^(#{1,6})([^\s#])", r"\1 \2", s, flags=re.M) + + # 문장 중간 헤더 분리: "...개요## 핵심" -> "...개요\n\n## 핵심" + s = re.sub(r"(? int(os.getenv("SUMMARY_LONGDOC_CHAR_LIMIT", "3500")) + + backend = os.getenv("SUMMARY_BACKEND", "hf").lower() + if not enable_long: + sys_txt = _system_prompt(dom, phase="final", output_format=output_format) + user_payload = _compose_user_payload(text, extra_context, output_format) # [CHANGED] + if backend == "ollama": + async for s in _stream_with_ollama(user_payload, system_text=sys_txt, output_format=output_format): + yield s + else: + async for s in _stream_with_hf(user_payload, system_text=sys_txt, output_format=output_format): + yield s + return + + # Long-doc: Map (chunk summaries) → Reduce (final synthesis streamed) + chunks = _chunk_text( + text, + chunk_chars=int(os.getenv("SUMMARY_CHUNK_CHARS", "2000")), + overlap=int(os.getenv("SUMMARY_CHUNK_OVERLAP", "200")), + ) + map_sys = _system_prompt(dom, phase="map", output_format=output_format) + partials: list[str] = [] + for idx, ch in enumerate(chunks, 1): + try: + map_input = _compose_user_payload(ch, "", output_format) # [CHANGED] + part = await _hf_generate_once(map_sys, map_input, max_new_tokens=int(os.getenv("HF_MAP_MAX_NEW_TOKENS", "220"))) + except Exception: + part = ch[:500] + partials.append(f"[Chunk {idx}]\n{part.strip()}") + + reduce_text = "\n\n".join(partials) + reduce_sys = _system_prompt(dom, phase="reduce", output_format=output_format) + reduce_input = _compose_user_payload(reduce_text, extra_context, output_format) # [CHANGED] + + if backend == "ollama": + async for s in _stream_with_ollama(reduce_input, system_text=reduce_sys, output_format=output_format): + yield s + else: + async for s in _stream_with_hf(reduce_input, system_text=reduce_sys, output_format=output_format): + yield s + + +# =============== 도메인/언어 감지 =============== +def _detect_domain(t: str) -> str: + s = (t or "").lower() + # code-like signals + if re.search(r"\b(def |class |import |#include|public\s+class|function\s|=>|:=)", s) or re.search(r"```|\bdiff --git\b|\bcommit\b", s): + return "code" + # paper-like signals + if re.search(r"\babstract\b|\bintroduction\b|\bmethod(s)?\b|\bresult(s)?\b|\bconclusion(s)?\b|doi:|arxiv:\d", s): + return "paper" + # meeting-like signals (KO/EN keywords) + if re.search(r"회의|안건|결정|논의|액션 아이템|참석자|회의록|meeting|agenda|minutes|action items|attendees", s): + return "meeting" + return "general" + +def _detect_lang(t: str) -> str: + """아주 단순한 언어 감지(영문자/한글자 수 비교). ko/en만 구분.""" + s = t or "" + en = len(re.findall(r"[A-Za-z]", s)) + ko = len(re.findall(r"[가-힣]", s)) + return "en" if en > ko else "ko" + + +# =============== 시스템 프롬프트 =============== +# [CHANGED] 출력 포맷(MD/HTML) 지원 + 마크다운 간격 규칙 + 도메인별 포함 요소 힌트 +def _system_prompt(domain: str, phase: str = "final", output_format: str = "md") -> str: + # phase: map | reduce | final + fmt = output_format.lower() + base_rules = ( + "역할: 너는 사실 보존에 강한 전문 요약가다. 입력 텍스트의 언어를 감지하고, 반드시 동일한 언어로 작성한다. " + "핵심 정보(주요 주장/결론, 인물·기관·수치·날짜·지표·범위, 원인↔결과·조건·한계)를 빠짐없이 담되 군더더기와 반복을 제거한다. " + "추정·가치판단·조언은 덧붙이지 않는다. 사고과정(Chain-of-Thought), 단계 나열, 메타 코멘트는 출력하지 않는다." + ) + if domain == "meeting": + include_hint = "결정 사항, 책임자/기한이 명시된 액션, 남은 이슈, 다음 단계" + elif domain == "code": + include_hint = "변경 목적/범위, 주요 API/모듈 영향, 호환성/리스크, 마이그레이션 포인트" + elif domain == "paper": + include_hint = "문제 정의·동기, 방법 요지, 데이터/세팅, 핵심 수치, 한계/가정" + else: + include_hint = "5W1H, 핵심 주장/결과, 중요한 수치·날짜·고유명사, 원인↔결과·조건, 한계/주의점" + + if fmt == "md": + format_rule = ( + "출력 형식: Markdown. 다음 섹션을 사용하라 — " + "# 제목, ## 개요, ## 핵심 요점(불릿 3–7개), ## 상세 설명(문단 분리), " + "## 용어 정리(필요시), ## 한계/주의, ## 할 일(있다면), ## 참고/추가자료(있다면). " + "각 섹션 제목만 출력하고, 불필요한 프리앰블은 쓰지 마라." + ) + else: + format_rule = ( + "출력 형식: HTML fragment.

,

,

,

,

    ,
  • , , 만 사용. " + "

    제목

    ,

    개요

    ,

    핵심 요점

      …
    ,

    상세 설명

    , " + "

    용어 정리

    ,

    한계/주의

    ,

    할 일

    ,

    참고/추가자료

    의 순서." + ) + + length_rule = "분량: 원문 대비 약 15–30%. 각 문단은 2–5문장." + # [CHANGED] 마크다운 간격 규칙 추가 + md_spacing_rule = ( + "마크다운 간격 규칙: 모든 헤더(#, ##, ### 등) 뒤에는 한 칸 공백을 두고, 헤더의 앞뒤에는 빈 줄 1줄을 둔다. " + "불릿(- )은 항목마다 줄바꿈하고, 서브항목은 들여쓰기 2–4칸을 사용한다." + ) + web_rule = ( + "추가자료가 주어지면 원문과 교차 검증하여 모순이 있으면 원문을 우선하라. " + "추가자료의 출처명(예: 위키백과)을 괄호로 간단히 표기할 수 있다." + ) + + if phase == "map": + scope = "이 청크만 대상으로 섹션 골격을 간략히 채워라. 과도한 요약 금지." + elif phase == "reduce": + scope = "아래 청크 요약들을 중복 없이 통합해 일관된 섹션 구성을 완성하라. 흐름(원인→과정→결과)을 유지." + else: + scope = "전체 텍스트를 위 섹션 구조에 맞춰 응집력 있게 작성하라. 첫 줄부터 본문만 출력." + + return f"{base_rules}\n포함 우선: {include_hint}\n{format_rule}\n{length_rule}\n{md_spacing_rule}\n{web_rule}\n{scope}" + + +# =============== HF backend (Transformers) =============== +_HF_MODEL = None +_HF_TOKENIZER = None +_HF_NAME = None + +def _resolve_dtype(): + if torch.cuda.is_available(): + if torch.cuda.is_bf16_supported(): + return torch.bfloat16 + return torch.float16 + return None + +def _load_hf_model(): + global _HF_MODEL, _HF_TOKENIZER, _HF_NAME + if _HF_MODEL is not None and _HF_TOKENIZER is not None: + return _HF_MODEL, _HF_TOKENIZER + + primary = os.getenv("HF_SUMMARY_MODEL", "CohereForAI/aya-23-8B") + fallback = os.getenv("HF_SUMMARY_FALLBACK_MODEL", "mistralai/Mistral-7B-Instruct-v0.3") + token = os.getenv("HF_API_TOKEN") or None + torch_dtype = _resolve_dtype() + load_in_4bit = os.getenv("HF_LOAD_IN_4BIT", "false").lower() in ("1", "true", "yes") + + def try_load(name: str): + hub_kwargs = {"trust_remote_code": True} + if token: + hub_kwargs["token"] = token + try: + tok = AutoTokenizer.from_pretrained(name, **hub_kwargs) + except Exception as e: + if "401" in str(e) or "Unauthorized" in str(e): + hk = dict(hub_kwargs); hk.pop("token", None) + tok = AutoTokenizer.from_pretrained(name, **hk) + else: + raise + kwargs = dict(hub_kwargs) + if load_in_4bit: + try: + kwargs.update({"load_in_4bit": True}) + except Exception: + pass + else: + if torch_dtype is not None: + kwargs["dtype"] = torch_dtype + def load_model(with_token: bool): + mk = dict(kwargs) + if not with_token and "token" in mk: + mk.pop("token", None) + return AutoModelForCausalLM.from_pretrained(name, device_map="auto", **mk) + + try: + model = load_model(with_token=bool(token)) + except Exception as e: + if "401" in str(e) or "Unauthorized" in str(e): + model = load_model(with_token=False) + else: + try: + model = AutoModelForCausalLM.from_pretrained(name, **kwargs) + except Exception: + mk = dict(kwargs); mk.pop("token", None) + model = AutoModelForCausalLM.from_pretrained(name, **mk) + if torch.cuda.is_available(): + try: model.to("cuda") + except Exception: pass + return model, tok + + try: + model, tok = try_load(primary) + _HF_MODEL, _HF_TOKENIZER, _HF_NAME = model, tok, primary + return _HF_MODEL, _HF_TOKENIZER + except Exception: + model, tok = try_load(fallback) + _HF_MODEL, _HF_TOKENIZER, _HF_NAME = model, tok, fallback + return _HF_MODEL, _HF_TOKENIZER + + +def _build_prompt(tokenizer, system_text: str, user_text: str) -> str: + messages = [ + {"role": "system", "content": system_text}, + {"role": "user", "content": user_text}, + ] + try: + return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) + except Exception: + # [CHANGED] 백업 프롬프트도 MD/HTML 구조 반영 + return ( + "You are a precise summarizer. Detect the input language (Korean/English) and write the summary in the SAME language. " + "Preserve key facts (claims, entities, numbers, dates); remove fluff; avoid speculation and chain-of-thought. " + "Use Markdown sections: # Title, ## Overview, ## Key Points, ## Details, ## Terms, ## Limitations, ## Action Items, ## References.\n\n" + "Text:\n" + user_text + "\n\nSummary:" + ) + + +async def _stream_with_hf(text: str, system_text: str | None = None, output_format: str = "md"): + model, tokenizer = _load_hf_model() + + sys_msg = system_text or ( + # [CHANGED] 기본 시스템 프롬프트: Markdown 섹션 + 동일 언어 + "역할: 너는 사실 보존에 강한 전문 요약가다. 입력 언어를 감지하고 동일 언어로 작성한다. " + "Markdown 섹션(# 제목, ## 개요, ## 핵심 요점, ## 상세 설명, ## 용어 정리, ## 한계/주의, ## 할 일, ## 참고/추가자료)을 사용한다. " + "핵심 주장/결과, 인물·기관·수치·날짜, 원인↔결과·조건·한계를 보존하고 군더더기는 제거한다. " + "추정·가치판단·조언 금지. 사고과정/단계 나열/메타 코멘트 금지. 각 문단 2–5문장." + ) + prompt = _build_prompt(tokenizer, sys_msg, text) + + streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True) + inputs = tokenizer(prompt, return_tensors="pt") + if torch.cuda.is_available(): + try: + inputs = {k: v.to(getattr(model, "device", "cuda")) for k, v in inputs.items()} + except Exception: + pass + + gen_kwargs = dict( + max_new_tokens=int(os.getenv("HF_MAX_NEW_TOKENS", "320")), # [CHANGED] 섹션 증가에 맞춰 상향 + do_sample=False, + repetition_penalty=1.05, + eos_token_id=tokenizer.eos_token_id, + pad_token_id=tokenizer.eos_token_id or tokenizer.pad_token_id, + streamer=streamer, + ) + if gen_kwargs.get("do_sample"): + gen_kwargs["temperature"] = float(os.getenv("HF_TEMPERATURE", "0.1")) + + def _gen(): + model.generate(**inputs, **gen_kwargs) + + thread = threading.Thread(target=_gen, daemon=True) + thread.start() + + q: asyncio.Queue = asyncio.Queue() + loop = asyncio.get_running_loop() + + def _drain_streamer(): + for chunk in streamer: + try: + asyncio.run_coroutine_threadsafe(q.put(chunk), loop) + except Exception: + break + try: + asyncio.run_coroutine_threadsafe(q.put(None), loop) + except Exception: + pass + + threading.Thread(target=_drain_streamer, daemon=True).start() + + buffer = "" + while True: + chunk = await q.get() + if chunk is None: + break + buffer += chunk + # [CHANGED] 플러시 트리거 확장 + 보정기 적용 + if buffer.endswith(("\n", "。", ".", "…", "!", "?", ")", "]")): + line = buffer # strip 하지 않음: 줄바꿈 유지 + buffer = "" + # [CHANGED] 마크다운 보정 + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + + # [CHANGED] 잔여 버퍼 마무리 보정 + if buffer.strip(): + line = buffer + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + + +async def _hf_generate_once(system_text: str, user_text: str, max_new_tokens: int = 256) -> str: + """Non-streaming single-shot generation (used for chunk map stage).""" + model, tokenizer = _load_hf_model() + prompt = _build_prompt(tokenizer, system_text, user_text) + + inputs = tokenizer(prompt, return_tensors="pt") + if torch.cuda.is_available(): + try: + inputs = {k: v.to(getattr(model, "device", "cuda")) for k, v in inputs.items()} + except Exception: + pass + + gen_kwargs = dict( + max_new_tokens=max_new_tokens, + do_sample=False, + repetition_penalty=1.05, + eos_token_id=tokenizer.eos_token_id, + pad_token_id=tokenizer.eos_token_id or tokenizer.pad_token_id, + ) + with torch.no_grad(): + out = model.generate(**inputs, **gen_kwargs) + gen_ids = out[0, inputs["input_ids"].shape[1]:] + return tokenizer.decode(gen_ids, skip_special_tokens=True).strip() + + +# =============== 텍스트 청크 나누기 =============== +def _chunk_text(text: str, chunk_chars: int = 2000, overlap: int = 200) -> list[str]: + text = text or "" + if len(text) <= chunk_chars: + return [text] + chunks = [] + start = 0 + while start < len(text): + end = start + chunk_chars + chunks.append(text[start:end]) + if end >= len(text): + break + start = max(end - overlap, start + 1) + return chunks + + +# =============== Ollama backend =============== +async def _stream_with_ollama(text: str, system_text: str | None = None, output_format: str = "md"): # 1) LangChain용 콜백 핸들러 cb = AsyncIteratorCallbackHandler() - # 2) Ollama Chat 모델 (streaming=True) - llm = ChatOllama( - base_url="http://localhost:11434", - model="qwen3:8b", - streaming=True, - callbacks=[cb], - temperature=0.6, - ) + # 2) Ollama Chat 모델 설정 (환경변수로 조정 가능) + base_url = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434") + primary_model = os.getenv("OLLAMA_MODEL", "qwen2.5:14b") + fallback_model = os.getenv("OLLAMA_FALLBACK_MODEL", "qwen2.5:7b") + num_ctx = os.getenv("OLLAMA_NUM_CTX") + num_thread = os.getenv("OLLAMA_NUM_THREAD") + + def make_llm(model_name: str) -> ChatOllama: + kwargs = {} + if num_ctx: + try: + kwargs["num_ctx"] = int(num_ctx) + except ValueError: + pass + if num_thread: + try: + kwargs["num_thread"] = int(num_thread) + except ValueError: + pass + return ChatOllama( + base_url=base_url, + model=model_name, + streaming=True, + callbacks=[cb], + temperature=0.4, + **kwargs, + ) - # 3) 프롬프트 + llm = make_llm(primary_model) + + # [CHANGED] Ollama 경로의 시스템 프롬프트도 MD 섹션 구조 지시 messages = [ SystemMessage( - content="다음 텍스트를 한국어로 간결하게 요약하세요. " - "사고 과정(Chain‑of‑Thought)은 절대 출력하지 마세요./no_think" + content=( + system_text or ( + "역할: 너는 사실 보존에 강한 전문 요약가다. 입력 언어를 감지하고 동일 언어로 작성한다. " + "Markdown 섹션(# 제목, ## 개요, ## 핵심 요점, ## 상세 설명, ## 용어 정리, ## 한계/주의, ## 할 일, ## 참고/추가자료)을 사용한다. " + "추정/가치판단/사고과정 금지. 각 문단 2–5문장. 마크다운 간격 규칙을 지켜라." + ) + ) ), HumanMessage(content=text), ] - # 4) LLM 호출 비동기 실행 - task = asyncio.create_task(llm.agenerate([messages])) + try: + task = asyncio.create_task(llm.agenerate([messages])) + except Exception: + cb = AsyncIteratorCallbackHandler() + llm = make_llm(fallback_model) + task = asyncio.create_task(llm.agenerate([messages])) buffer = "" async for token in cb.aiter(): buffer += token - if buffer.endswith(("\n", "。", ".", "…")): - line = buffer.strip() + # [CHANGED] 플러시 트리거 확장 + 마크다운 보정 적용 + if buffer.endswith(("\n", "。", ".", "…", "!", "?", ")", "]")): + line = buffer buffer = "" + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + + # [CHANGED] 잔여 버퍼 마무리 보정 + if buffer.strip(): + line = buffer + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + + try: + await task + except Exception as e: + if "not found" in str(e).lower() or "model \"" in str(e).lower(): + cb2 = AsyncIteratorCallbackHandler() + llm2 = make_llm(fallback_model) + task2 = asyncio.create_task(llm2.agenerate([messages])) + buffer2 = "" + async for token in cb2.aiter(): + buffer2 += token + if buffer2.endswith(("\n", "。", ".", "…", "!", "?", ")", "]")): + line = buffer2 + buffer2 = "" + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + # [CHANGED] 잔여 버퍼 + if buffer2.strip(): + line = buffer2 + if output_format.lower() == "md": + line = _format_md_stream(line) + if not _THOUGHT_PAT.match(line.strip()): + yield f"data: {line}\n\n" + await task2 + else: + raise + + +# =============== 보조 유틸리티 =============== +# [CHANGED] 원문 + (선택) 추가자료를 모델에 전달하기 위한 합본 +def _compose_user_payload(main_text: str, extra_context: str, output_format: str) -> str: + fmt = output_format.lower() + if fmt == "md": + if extra_context: + return ( + "## 원문\n" + f"{main_text}\n\n" + "## 추가자료(요약)\n" + f"{extra_context}\n" + ) + return main_text + else: + if extra_context: + return f"

    원문

    \n{main_text}\n\n

    추가자료(요약)

    \n{extra_context}\n" + return main_text - if not _THOUGHT_PAT.match(line): - yield f"data: {line}\n\n" # SSE 청크 전송 +def _is_augmentation_allowed() -> bool: + """환경변수로 보강 ON/OFF 제어. 기본 False.""" + return os.getenv("AUGMENT_WEB", "false").lower() in ("1", "true", "yes") - await task # 예외 전파 +def _extract_entities_for_web(text: str, lang: str = "ko", max_items: int = 5) -> list[str]: + """ + 매우 가벼운 엔티티 후보 추출: + - 영문: 대문자로 시작하는 2~4단어 구 + - 한글: 괄호/따옴표 내 주요어 + 2자 이상 단어 + """ + items: list[str] = [] + if lang == "en": + items = re.findall(r"\b([A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,3})\b", text) + else: + items = re.findall(r"[「『“\"'(](.*?)[」』”\"')]", text) + re.findall(r"[가-힣A-Za-z][가-힣A-Za-z]{1,}", text) + stop = set(["회의", "안건", "결정", "논의", "데이터", "모델", "결과", "방법", "프로젝트", "사용자", "시스템"]) + uniq = [] + for w in items: + w = re.sub(r"[\s]+", " ", w).strip() + if 2 <= len(w) <= 80 and w not in stop and not re.match(r"^\d+$", w): + if w not in uniq: + uniq.append(w) + return uniq[:max_items] + +def _fetch_wikipedia_summaries(entities: list[str], lang: str = "ko", max_sources: int = 5) -> str: + """ + 간단한 Wikipedia summary 수집. + - 공개 REST 엔드포인트 사용(무인증) + - 실패/타임아웃은 건너뜀 + - 결과는 Markdown 불릿로 반환 + """ + base = "https://%s.wikipedia.org/api/rest_v1/page/summary/%s" % (("ko" if lang == "ko" else "en"), "%s") + out = [] + timeout = float(os.getenv("AUGMENT_HTTP_TIMEOUT", "2.5")) + session = requests.Session() + headers = {"User-Agent": os.getenv("AUGMENT_UA", "SummaryAgent/1.0")} + for ent in entities[:max_sources]: + url = base % _urlquote(ent) + try: + r = session.get(url, headers=headers, timeout=timeout) + if r.status_code != 200: + continue + data = r.json() + title = data.get("title") or ent + extract = (data.get("extract") or "").strip() + if not extract: + continue + extract = (extract[:500] + "…") if len(extract) > 500 else extract + # ko/en 모두 동일한 표기 + src = "위키백과" if lang == "ko" else "Wikipedia" + out.append(f"- **{title}** ({src}): {extract}") + time.sleep(0.05) + except Exception: + continue + return "\n".join(out)