feat(screener): ai_news Claude Haiku analyzer (-10~+10 + clamp + JSON-fail soft)

stock-lab/app/screener/ai_news/analyzer.py

@@ -0,0 +1,76 @@
+"""Claude Haiku 기반 종목 뉴스 호재/악재 분석."""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+from typing import Any, Dict, List
+
+log = logging.getLogger(__name__)
+
+DEFAULT_MODEL = os.getenv("AI_NEWS_MODEL", "claude-haiku-4-5-20251001")
+
+PROMPT_TEMPLATE = """다음은 종목 {name}({ticker})에 대한 최근 뉴스 {n}개의 헤드라인입니다.
+
+{news_block}
+
+이 뉴스들이 종목에 호재인지 악재인지 평가하세요.
+score: -10(매우 강한 악재) ~ +10(매우 강한 호재) 사이의 실수. 0은 중립.
+reason: 30자 이내 한 줄 근거.
+
+JSON으로만 응답하세요. 다른 텍스트 금지:
+{{"score": <float>, "reason": "<string>"}}"""
+
+
+def _clamp(x: float, lo: float = -10.0, hi: float = 10.0) -> float:
+    return max(lo, min(hi, x))
+
+
+async def score_sentiment(
+    llm,
+    ticker: str,
+    news: List[Dict[str, Any]],
+    *,
+    name: str | None = None,
+    model: str = DEFAULT_MODEL,
+) -> Dict[str, Any]:
+    """Returns {ticker, score_raw, reason, news_count, tokens_input, tokens_output, model}."""
+    news_block = "\n".join(f"- {n['title']}" for n in news)
+    prompt = PROMPT_TEMPLATE.format(
+        name=name or ticker, ticker=ticker,
+        n=len(news), news_block=news_block,
+    )
+    resp = await llm.messages.create(
+        model=model,
+        max_tokens=200,
+        messages=[{"role": "user", "content": prompt}],
+    )
+    text = resp.content[0].text if resp.content else ""
+    in_tokens = int(getattr(resp.usage, "input_tokens", 0) or 0)
+    out_tokens = int(getattr(resp.usage, "output_tokens", 0) or 0)
+
+    try:
+        data = json.loads(text)
+        score = _clamp(float(data["score"]))
+        reason = str(data["reason"])[:200]
+        return {
+            "ticker": ticker,
+            "score_raw": score,
+            "reason": reason,
+            "news_count": len(news),
+            "tokens_input": in_tokens,
+            "tokens_output": out_tokens,
+            "model": model,
+        }
+    except (json.JSONDecodeError, KeyError, TypeError, ValueError) as e:
+        log.warning("ai_news parse fail for %s: %s (raw=%r)", ticker, e, text[:100])
+        return {
+            "ticker": ticker,
+            "score_raw": 0.0,
+            "reason": f"parse fail: {e!s}"[:200],
+            "news_count": len(news),
+            "tokens_input": in_tokens,
+            "tokens_output": out_tokens,
+            "model": model,
+        }

stock-lab/tests/test_ai_news_analyzer.py

@@ -0,0 +1,55 @@
+import json
+import pytest
+from unittest.mock import AsyncMock, MagicMock
+
+from app.screener.ai_news import analyzer
+
+
+def _mk_llm(content_text: str, in_tokens: int = 100, out_tokens: int = 20):
+    llm = AsyncMock()
+    resp = MagicMock()
+    block = MagicMock()
+    block.text = content_text
+    resp.content = [block]
+    resp.usage = MagicMock(input_tokens=in_tokens, output_tokens=out_tokens)
+    llm.messages = MagicMock()
+    llm.messages.create = AsyncMock(return_value=resp)
+    return llm
+
+
+NEWS = [{"title": "삼성전자, HBM 양산"}, {"title": "메모리 가격 반등"}]
+
+
+@pytest.mark.asyncio
+async def test_score_sentiment_success_parses_json():
+    llm = _mk_llm(json.dumps({"score": 7.5, "reason": "HBM 호재"}))
+    out = await analyzer.score_sentiment(llm, "005930", NEWS, name="삼성전자")
+    assert out["ticker"] == "005930"
+    assert out["score_raw"] == 7.5
+    assert out["reason"] == "HBM 호재"
+    assert out["news_count"] == 2
+    assert out["tokens_input"] == 100
+    assert out["tokens_output"] == 20
+
+
+@pytest.mark.asyncio
+async def test_score_sentiment_json_parse_fail_returns_zero():
+    llm = _mk_llm("not valid json")
+    out = await analyzer.score_sentiment(llm, "005930", NEWS)
+    assert out["score_raw"] == 0.0
+    assert "parse fail" in out["reason"]
+    assert out["tokens_input"] == 100  # 호출은 발생했음
+
+
+@pytest.mark.asyncio
+async def test_score_sentiment_clamps_out_of_range():
+    llm = _mk_llm(json.dumps({"score": 15.0, "reason": "초강세"}))
+    out = await analyzer.score_sentiment(llm, "005930", NEWS)
+    assert out["score_raw"] == 10.0  # +10 클램프
+
+
+@pytest.mark.asyncio
+async def test_score_sentiment_clamps_negative_out_of_range():
+    llm = _mk_llm(json.dumps({"score": -42.0, "reason": "초악재"}))
+    out = await analyzer.score_sentiment(llm, "005930", NEWS)
+    assert out["score_raw"] == -10.0