web-page-backend/stock/app/ai_summarizer.py

"""LLM 기반 뉴스 요약 모듈.

LLM_PROVIDER 환경변수로 provider 전환:
  - claude (기본): Anthropic Messages API (claude-haiku-4-5)
  - ollama: Windows AI 서버의 Ollama (qwen3:14b 등)

`summarize_news(articles)` 시그니처는 provider와 무관하게 동일하며,
실패 시 `LLMError`(구 `OllamaError` alias)를 raise 한다.
"""
import os
import logging
import time
from typing import List, Dict, Any

import httpx

logger = logging.getLogger("stock.ai_summarizer")

LLM_PROVIDER = os.getenv("LLM_PROVIDER", "claude").lower().strip()

# Ollama
OLLAMA_URL = os.getenv("OLLAMA_URL", "http://192.168.45.59:11435")
OLLAMA_MODEL = os.getenv("OLLAMA_MODEL", "qwen3:14b")

# Anthropic (Claude)
ANTHROPIC_API_KEY = os.getenv("ANTHROPIC_API_KEY", "")
ANTHROPIC_MODEL = os.getenv("ANTHROPIC_MODEL", "claude-haiku-4-5-20251001")
ANTHROPIC_URL = "https://api.anthropic.com/v1/messages"
ANTHROPIC_VERSION = "2023-06-01"

_PROMPT_TEMPLATE = """당신은 한국 주식 시장 애널리스트입니다. 아래 뉴스 목록을 읽고 투자자 관점에서 한국어로 간결하게 요약하세요.

반드시 아래 형식을 그대로 지켜서 출력하세요. 다른 설명이나 서두, `<think>` 같은 태그는 절대 출력하지 마세요.

📌 시장 흐름
(2줄 요약)

🔥 주목 이슈
• (이슈 1)
• (이슈 2)
• (이슈 3)

💡 투자 관점
(1줄 인사이트)

=== 뉴스 목록 ===
{news_block}
"""


class LLMError(RuntimeError):
    """LLM provider 호출 실패."""


# 하위 호환 alias (main.py 등 기존 import 유지)
OllamaError = LLMError


def _build_news_block(articles: List[Dict[str, Any]]) -> str:
    lines = []
    for i, art in enumerate(articles, start=1):
        title = (art.get("title") or "").strip()
        content = (art.get("content") or art.get("summary") or "").strip()
        if content:
            lines.append(f"{i}. {title} — {content}")
        else:
            lines.append(f"{i}. {title}")
    return "\n".join(lines) if lines else "(뉴스 없음)"


async def _summarize_with_ollama(prompt: str) -> Dict[str, Any]:
    url = f"{OLLAMA_URL.rstrip('/')}/api/generate"
    payload = {"model": OLLAMA_MODEL, "prompt": prompt, "stream": False}

    started = time.monotonic()
    try:
        async with httpx.AsyncClient(timeout=180.0) as client:
            resp = await client.post(url, json=payload)
    except httpx.HTTPError as e:
        err_type = type(e).__name__
        err_msg = str(e) or "(no message)"
        logger.error(f"Ollama 연결 실패 ({url}): [{err_type}] {err_msg}")
        raise LLMError(f"Ollama 연결 실패: [{err_type}] {err_msg}") from e

    if resp.status_code != 200:
        logger.error(f"Ollama 응답 오류 {resp.status_code}: {resp.text[:200]}")
        raise LLMError(f"Ollama HTTP {resp.status_code}: {resp.text[:200]}")

    try:
        data = resp.json()
    except ValueError as e:
        raise LLMError(f"Ollama 응답 JSON 파싱 실패: {e}") from e

    summary = (data.get("response") or "").strip()
    prompt_tokens = int(data.get("prompt_eval_count") or 0)
    completion_tokens = int(data.get("eval_count") or 0)
    total_duration_ns = int(data.get("total_duration") or 0)
    if total_duration_ns > 0:
        duration_ms = total_duration_ns // 1_000_000
    else:
        duration_ms = int((time.monotonic() - started) * 1000)

    return {
        "summary": summary,
        "tokens": {
            "prompt": prompt_tokens,
            "completion": completion_tokens,
            "total": prompt_tokens + completion_tokens,
        },
        "model": data.get("model") or OLLAMA_MODEL,
        "duration_ms": duration_ms,
    }


async def _summarize_with_claude(prompt: str) -> Dict[str, Any]:
    if not ANTHROPIC_API_KEY:
        raise LLMError("ANTHROPIC_API_KEY 미설정 — Claude provider 사용 불가")

    headers = {
        "x-api-key": ANTHROPIC_API_KEY,
        "anthropic-version": ANTHROPIC_VERSION,
        "content-type": "application/json",
    }
    payload = {
        "model": ANTHROPIC_MODEL,
        "max_tokens": 1024,
        "messages": [{"role": "user", "content": prompt}],
    }

    started = time.monotonic()
    try:
        async with httpx.AsyncClient(timeout=60.0) as client:
            resp = await client.post(ANTHROPIC_URL, headers=headers, json=payload)
    except httpx.HTTPError as e:
        err_type = type(e).__name__
        err_msg = str(e) or "(no message)"
        logger.error(f"Anthropic 연결 실패: [{err_type}] {err_msg}")
        raise LLMError(f"Anthropic 연결 실패: [{err_type}] {err_msg}") from e

    if resp.status_code != 200:
        logger.error(f"Anthropic 응답 오류 {resp.status_code}: {resp.text[:300]}")
        raise LLMError(f"Anthropic HTTP {resp.status_code}: {resp.text[:200]}")

    try:
        data = resp.json()
    except ValueError as e:
        raise LLMError(f"Anthropic 응답 JSON 파싱 실패: {e}") from e

    # content: [{"type": "text", "text": "..."}]
    blocks = data.get("content") or []
    summary = "".join(b.get("text", "") for b in blocks if b.get("type") == "text").strip()

    usage = data.get("usage") or {}
    prompt_tokens = int(usage.get("input_tokens") or 0)
    completion_tokens = int(usage.get("output_tokens") or 0)
    duration_ms = int((time.monotonic() - started) * 1000)

    return {
        "summary": summary,
        "tokens": {
            "prompt": prompt_tokens,
            "completion": completion_tokens,
            "total": prompt_tokens + completion_tokens,
        },
        "model": data.get("model") or ANTHROPIC_MODEL,
        "duration_ms": duration_ms,
    }


async def summarize_news(articles: List[Dict[str, Any]]) -> Dict[str, Any]:
    """뉴스 리스트를 LLM으로 요약. provider는 LLM_PROVIDER 환경변수로 선택.

    Returns:
        {"summary": str, "tokens": {...}, "model": str, "duration_ms": int}
    Raises:
        LLMError: provider 호출 실패 시.
    """
    prompt = _PROMPT_TEMPLATE.format(news_block=_build_news_block(articles))

    if LLM_PROVIDER == "ollama":
        return await _summarize_with_ollama(prompt)
    if LLM_PROVIDER == "claude":
        return await _summarize_with_claude(prompt)
    raise LLMError(f"지원하지 않는 LLM_PROVIDER: {LLM_PROVIDER}")