GPT — 智能问答系统

GPT/
├── qa_system/
│   ├── app.py            # Flask 入口 + 日志配置
│   ├── database.py       # 数据库连接与问答对 CRUD
│   ├── nlp_processor.py  # 嵌入向量 + 相似度检索
│   ├── init_data.py      # 初始化数据
│   └── templates/
├── models/               # 模型缓存目录
├── logs/                 # 日志
├── knowledge.db          # SQLite 知识库
└── README.md

def __init__(self, model_name='paraphrase-multilingual-MiniLM-L12-v2'):
    cache_dir = Path('models')
    cache_dir.mkdir(exist_ok=True)
    os.environ['TRANSFORMERS_CACHE'] = str(cache_dir)
    os.environ['HF_HOME'] = str(cache_dir)
    self.model = SentenceTransformer(model_name, cache_folder=str(cache_dir))

def _get_simple_embedding(self, text: str) -> np.ndarray:
    """简单的文本向量化方法（100维哈希词袋）"""
    words = text.lower().split()
    vector = np.zeros(100)
    for i, word in enumerate(words):
        vector[hash(word) % 100] += 1     # 哈希到 100 维桶
    norm = np.linalg.norm(vector)
    if norm > 0:
        vector = vector / norm            # L2 归一化
    return vector

def calculate_similarity(self, embedding1, embedding2) -> float:
    return np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))

def find_most_similar(self, query, candidates, threshold=0.5):
    query_embedding = self.get_embedding(query)
    max_similarity = 0
    best_answer = None
    for row in candidates:
        id_, question, answer, embedding_str, created_at = row
        embedding = np.array([float(x) for x in embedding_str.strip('[]').split(',')])
        similarity = self.calculate_similarity(query_embedding, embedding)
        if similarity > max_similarity:
            max_similarity = similarity
            best_answer = answer
    if max_similarity < threshold:
        return None, 0        # 低于阈值判定为"无答案"
    return best_answer, max_similarity

用户输入 → AJAX POST /api/ask → 数据库取全部问答对
        → 问题转嵌入向量 → 逐一算余弦相似度
        → 取最高分且 > 0.5 → 返回答案（否则返回"无匹配"）