AI 예측 입력 통일·데이터 전면 보강

- 3모델 페르소나 분화 제거 — 동일 프롬프트+동일 데이터 (차이는 모델 판단만)
- 순위 라인에 게임차·팀 타율·팀 ERA 추가
- 전날 결과+활약(홈런·맹타), 확정 라인업+타자 시즌 타율 주입
  (응원가 파이프라인 헬퍼 재사용, 라인업 미발표 시 생략)
- 선발 상대 ERA 0.00(기록 없음) 오표기 생략

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

backend/app/services/ai.py

@@ -34,46 +34,21 @@ class MatchContext:
     league: str = "wc"  # wc(축구) | kbo | mlb | mls — 프롬프트·스코어 범위 분기
 
 
-# 모델별 분석 관점(페르소나) — 동일 경기라도 서로 다른 시각으로 보게 해
+# 공통 페르소나 — 세 모델 모두 동일한 입력(동일 프롬프트+동일 데이터)을 받는다.
-# 예측이 자연스럽게 갈리도록 한다(강제 분산이 아니라 진짜 판단의 다양화).
+# 예측 차이는 모델 자체의 판단 차이에서만 나온다.
-PERSONA = {
+ANALYST = (
-    "GPT": (
+    "You are an expert soccer analyst. Use ALL of the factual match data provided "
-        "You are a DATA-DRIVEN analyst. Base your call on recent form, head-to-head, "
+    "below — recent form, standings, head-to-head and matchup context — and weigh "
-        "FIFA ranking gaps and goals-scored/conceded trends. Be objective and "
+    "it over your prior knowledge to make the most accurate prediction possible."
-        "evidence-led; pick the scoreline the numbers most support."
+)
-    ),
-    "Claude": (
-        "You are a TACTICAL analyst. Focus on matchups, defensive organization, "
-        "midfield control and game-state. Take low-scoring games, tight margins and "
-        "genuine upset potential seriously — do not just rubber-stamp the favorite."
-    ),
-    "Gemini": (
-        "You are an ATTACKING-MINDED analyst. Weigh momentum, attacking quality, "
-        "star players and scoring potential. Lean toward open, higher-scoring "
-        "scenarios when the talent and tempo justify it."
-    ),
-}
 
-
+BASEBALL_ANALYST = (
-# 야구용 페르소나 — 축구 페르소나와 같은 3분화(데이터/수비·투수/공격) 구도.
+    "You are an expert baseball analyst. Use ALL of the factual match data provided "
-BASEBALL_PERSONA = {
+    "below — recent form, standings with games-behind, team batting/ERA, "
-    "GPT": (
+    "head-to-head record, starting pitchers, confirmed lineups and yesterday's "
-        "You are a DATA-DRIVEN baseball analyst. Base your call on recent form, "
+    "results — and weigh it over your prior knowledge to make the most accurate "
-        "season standings, head-to-head record and run-scored/allowed trends. "
+    "prediction possible."
-        "Be objective and evidence-led; pick the scoreline the numbers most support."
+)
-    ),
-    "Claude": (
-        "You are a PITCHING-AND-DEFENSE analyst. Weigh starting rotation strength, "
-        "bullpen fatigue, and defensive quality. Take low-scoring games, tight "
-        "margins and genuine upset potential seriously — do not just rubber-stamp "
-        "the favorite."
-    ),
-    "Gemini": (
-        "You are an OFFENSE-MINDED analyst. Weigh lineup depth, power hitting, "
-        "momentum and ballpark factors. Lean toward open, higher-scoring scenarios "
-        "when the bats and conditions justify it."
-    ),
-}
 
 _LEAGUE_LABEL = {
     "kbo": "2026 KBO League (Korean professional baseball) regular-season game",
@@ -82,7 +57,7 @@ _LEAGUE_LABEL = {
 
 
 def _prompt_baseball(ctx: MatchContext, model: str) -> str:
-    persona = BASEBALL_PERSONA[model]
+    persona = BASEBALL_ANALYST  # 모델 공통 — model 파라미터는 호출부 호환용
     data = f"\n{ctx.data_block}\n" if ctx.data_block else ""
     draw_note = (
         "KBO regular-season games can end in a DRAW after 12 innings, but draws "
@@ -92,9 +67,7 @@ def _prompt_baseball(ctx: MatchContext, model: str) -> str:
     )
     return (
         f"{persona}\n"
-        f"Predict the result of this {_LEAGUE_LABEL[ctx.league]} using YOUR "
+        f"Predict the result of this {_LEAGUE_LABEL[ctx.league]}.\n"
-        f"perspective above. Judge independently — it is fine to differ from the "
-        f"obvious consensus pick when your perspective warrants it.\n"
         f"Team A (away): {ctx.team_a}\nTeam B (home): {ctx.team_b}\n"
         f"Ballpark: {ctx.venue}\nFirst pitch: {ctx.kickoff}\n"
         f"{data}"
@@ -115,9 +88,7 @@ def _prompt(ctx: MatchContext, persona: str) -> str:
     data = f"\n{ctx.data_block}\n" if ctx.data_block else ""
     return (
         f"{persona}\n"
-        f"Predict the result of this 2026 FIFA World Cup match using YOUR perspective "
+        f"Predict the result of this 2026 FIFA World Cup match.\n"
-        f"above. Judge independently — it is fine to differ from the obvious consensus "
-        f"pick when your perspective warrants it.\n"
         f"Team A: {ctx.team_a}\nTeam B: {ctx.team_b}\n"
         f"Venue: {ctx.venue}\nKickoff: {ctx.kickoff}\n"
         f"{data}"
@@ -142,8 +113,7 @@ def _prompt_mls(ctx: MatchContext, persona: str) -> str:
     return (
         f"{persona}\n"
         f"Predict the result of this 2026 MLS (Major League Soccer) regular-season "
-        f"match using YOUR perspective above. Judge independently — it is fine to "
+        f"match.\n"
-        f"differ from the obvious consensus pick when your perspective warrants it.\n"
         f"Team A (away): {ctx.team_a}\nTeam B (home): {ctx.team_b}\n"
         f"Venue: {ctx.venue}\nKickoff: {ctx.kickoff}\n"
         f"{data}"
@@ -166,8 +136,8 @@ def _build_prompt(ctx: MatchContext, model: str) -> str:
     if ctx.league in ("kbo", "mlb"):
         return _prompt_baseball(ctx, model)
     if ctx.league == "mls":
-        return _prompt_mls(ctx, PERSONA[model])
+        return _prompt_mls(ctx, ANALYST)
-    return _prompt(ctx, PERSONA[model])
+    return _prompt(ctx, ANALYST)
 
 
 # JSON Schema (구조화 출력용 — Anthropic/OpenAI 공통)

backend/app/services/baseball_data.py

@@ -89,17 +89,26 @@ def _starter_line(side: str, s: dict | None) -> str | None:
         f" ({s['w']}W-{s['l']}L)"
         if s.get("w") is not None and s.get("l") is not None else ""
     )
-    vs = f", ERA vs this opponent {s['vsEra']}" if s.get("vsEra") else ""
+    # 상대 ERA 0.00 은 대부분 '대전 기록 없음' — 무실점으로 오해하지 않게 생략
+    vs_val = s.get("vsEra")
+    vs = (
+        f", ERA vs this opponent {vs_val}"
+        if vs_val not in (None, "", 0, "0", "0.00", "-")
+        else ""
+    )
     return f"[{side} starting pitcher] {s['name']}{era}{rec}{vs}"
 
 
 def _standing_line(name: str, st: dict | None) -> str | None:
     if not st:
         return None
+    gb = f", GB {st['gb']}" if st.get("gb") not in (None, "", 0, "0.0") else ""
+    ba = f", team AVG {st['avg']}" if st.get("avg") else ""
+    era = f", team ERA {st['era']}" if st.get("era") else ""
     extra = f", last5 {st['last5']}" if st.get("last5") else ""
     return (
         f"[{name} standings] rank {st.get('rank')}, {st.get('w')}W-"
-        f"{st.get('l')}L (pct {st.get('wra')}){extra}"
+        f"{st.get('l')}L (pct {st.get('wra')}){gb}{ba}{era}{extra}"
     )
 
 
@@ -132,6 +141,40 @@ async def _cached_extras(db, match: Match) -> list[str]:
     return lines
 
 
+async def _context_extras(db, match: Match) -> list[str]:
+    """전날 결과·활약 + 확정 라인업(타자 시즌 타율) — 응원가 파이프라인 헬퍼 재사용.
+
+    라인업은 발표 전이면 생략(예측이 킥오프 22h 전부터 생성되므로 보통 미포함).
+    """
+    from .songs import _lineup_line, _yesterday_info, fetch_lineups
+
+    lines: list[str] = []
+    avg_maps: dict[str, dict] = {}
+    for side, code, name in (
+        ("a", match.team_a_code, match.team_a_short),
+        ("b", match.team_b_code, match.team_b_short),
+    ):
+        label = "Away" if side == "a" else "Home"
+        try:
+            y, avg_map = await _yesterday_info(db, match, code, name)
+        except Exception:  # noqa: BLE001 — 부가 데이터 실패는 생략
+            y, avg_map = None, {}
+        avg_maps[side] = avg_map
+        if y:
+            lines.append(f"[{label} yesterday] {y}")
+    try:
+        lu = await fetch_lineups(match)
+    except Exception:  # noqa: BLE001
+        lu = None
+    if lu and lu.get("announced"):
+        for side in ("a", "b"):
+            label = "Away" if side == "a" else "Home"
+            line = _lineup_line(lu, side, avg_maps.get(side))
+            if line:
+                lines.append(f"[{label}] {line}")
+    return lines
+
+
 async def build_baseball_data_block(db, match: Match) -> str | None:
     ga = await _team_games(db, match.league, match.team_a_code, match.kickoff_at)
     gb = await _team_games(db, match.league, match.team_b_code, match.kickoff_at)
@@ -141,11 +184,13 @@ async def build_baseball_data_block(db, match: Match) -> str | None:
     h2h = await _h2h_line(
         db, match.league, match.team_a_code, match.team_b_code, match.kickoff_at
     )
+    context = await _context_extras(db, match)
     return "\n".join([
         "=== MATCH DATA (factual; weigh heavily over priors) ===",
         "Away " + _team_lines(match.team_a_name, ga),
         "Home " + _team_lines(match.team_b_name, gb),
         f"[Head-to-head in our data] {h2h}",
         *extras,
+        *context,
         "===",
     ])