app.py

from fastapi import FastAPI
from pydantic import BaseModel
from sentence_transformers import CrossEncoder
import fugashi
import numpy as np
from fastapi.middleware.cors import CORSMiddleware

app = FastAPI()

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"], 
    allow_methods=["*"],
    allow_headers=["*"],
)

# モデルのロード（起動時に1回だけ）
nli_model = CrossEncoder("akiFQC/bert-base-japanese-v3_nli-jsnli-jnli-jsick")
tagger = fugashi.Tagger()

# --- 新規追加: 判定用の定数と辞書 ---
SIMILARITY_THRESHOLD = 0.85

ALLOWED_POS_GROUPS = [
    {"名詞"},
    {"形容詞", "形容動詞", "形状詞"},
    {"動詞", "名詞"},
    {"副詞", "名詞"},
]

HONORIFIC_PREFIXES = {"お", "ご", "御"}

# 形状詞も有効な品詞として追加
VALID_POS = {"名詞", "動詞", "形容詞", "形容動詞", "副詞", "接頭辞", "形状詞"}

CUSTOM_SYNONYMS = [
    {"スマホ", "スマートフォン"},
    {"チャリ", "自転車"},
    {"おでこ", "額"},
    {"ビリ", "最下位"},
    {"おにぎり", "おむすび"},
    {"めっちゃ", "とても"}
]

class StringInput(BaseModel):
    user_input: str
    correct_word: str

def is_valid_japanese(word: str) -> tuple[bool, str]:
    if not word.strip():
        return False, "入力が空です"
    if len(word.strip()) > 10:
        return False, "10文字以内で入力してください"
    tokens = list(tagger(word))
    if not tokens:
        return False, "日本語として認識できません"
    has_valid_token = any(t.feature.pos1 in VALID_POS for t in tokens)
    if not has_valid_token:
        return False, "単語として成立していません"
    return True, "OK"

def get_word_info(word: str) -> dict:
    tokens = list(tagger(word))
    pos = "不明"
    for token in tokens:
        # お・ご・御 はスキップ
        if token.feature.pos1 == "接頭辞" and token.surface in HONORIFIC_PREFIXES:
            continue
        # 助詞や助動詞など、単語のメインにならない付属語は無視して上書きしない
        if token.feature.pos1 in ["助詞", "助動詞", "補助記号"]:
            continue
        # 複合語対策: ループを最後まで回して一番最後の主要な品詞を取得する
        pos = token.feature.pos1
        
    surfaces = [t.surface for t in tokens]
    last_surface = surfaces[-1] if surfaces else ""
    
    if last_surface in ["た", "だ"]:
        tense = "過去"
    elif last_surface in ["いる", "いた", "います"]:
        tense = "進行・状態"
    elif last_surface in ["る", "す", "く", "ぬ"]:
        tense = "現在"
    else:
        tense = "その他"
    
    return {"pos": pos, "tense": tense, "surfaces": surfaces}

def pos_compatible(pos1: str, pos2: str) -> bool:
    if pos1 == pos2:
        return True
    for group in ALLOWED_POS_GROUPS:
        if pos1 in group and pos2 in group:
            return True
    return False

# --- 新規追加: Softmax関数 ---
def softmax(x):
    e_x = np.exp(x - np.max(x))
    return e_x / e_x.sum()

def check_answer(user_input: str, correct_word: str) -> dict:
    # 0. バリデーションチェック
    valid, reason = is_valid_japanese(user_input)
    if not valid:
        return {"result": "invalid", "reason": reason, "score": 0.0}

    # --- 新規追加: 完全一致チェック（AI判定や品詞チェックをすべてバイパス） ---
    # .strip() を使って、目に見えない前後のスペースや改行を無視して比較します
    if user_input.strip() == correct_word.strip():
        return {"result": "正解", "reason": "完全一致", "score": 1.0}
    # ------------------------------------------------------------------

    # 1. カスタム辞書による絶対正解のチェック（AI判定をバイパス）
    for syn_set in CUSTOM_SYNONYMS:
        if user_input in syn_set and correct_word in syn_set:
            return {"result": "正解", "reason": "同義語（辞書一致）", "score": 1.0}

    user_info = get_word_info(user_input)
    correct_info = get_word_info(correct_word)

    # 2. 品詞チェック
    if not pos_compatible(user_info["pos"], correct_info["pos"]):
        return {
            "result": "不正解",
            "reason": f"品詞が違います（{user_info['pos']} vs {correct_info['pos']}）",
            "score": 0.0
        }

    # 3. 時制チェック
    if user_info["pos"] == "動詞" and correct_info["pos"] == "動詞":
        if user_info["tense"] != correct_info["tense"]:
            return {
                "result": "不正解",
                "reason": f"時制が違います（{user_info['tense']} vs {correct_info['tense']}）",
                "score": 0.0
            }

    # 4. AI (NLIモデル) によるスコア算出と Softmax 変換
    labels = ["含意", "中立", "矛盾"]
    scores_ab = nli_model.predict([(user_input, correct_word)])[0]
    scores_ba = nli_model.predict([(correct_word, user_input)])[0]

    res_ab = np.argmax(scores_ab)
    res_ba = np.argmax(scores_ba)
    
    probs_ab = softmax(scores_ab)
    probs_ba = softmax(scores_ba)

    # 5. 最終判定ロジック
    if res_ab == 0 and res_ba == 0:
        avg_prob = (probs_ab[0] + probs_ba[0]) / 2
        if avg_prob <= SIMILARITY_THRESHOLD:
            return {"result": "不正解", "reason": "似ているが違う", "score": 0.0}
        return {"result": "正解", "reason": "同義語", "score": 1.0}
        
    elif res_ab == 0 or res_ba == 0:
        return {"result": "不正解", "reason": "包含関係（上位/下位概念）", "score": 0.5}
    else:
        return {"result": "不正解", "reason": labels[res_ab], "score": 0.0}

@app.get("/")
def root():
    return {"status": "ok"}

@app.post("/check")
def check(data: StringInput):
    # APIリクエスト・レスポンスの確認用ログは保持
    print(f"受信: user_input='{data.user_input}' / correct_word='{data.correct_word}'")
    result = check_answer(data.user_input, data.correct_word)
    print(f"結果: {result}")
    return result