Update api.py

api.py

@@ -203,77 +203,6 @@ def generate_text_sample(model, prompt, max_len=100, max_gen=98,
         decoded = decoded.replace(t, "")
     return decoded.strip()
 
-
-from sklearn.feature_extraction.text import TfidfVectorizer
-from sklearn.decomposition import TruncatedSVD
-from sklearn.metrics.pairwise import cosine_similarity
-
-class SimilarityMemory:
-    def __init__(self, n_components=100):
-        self.memory_texts = []
-        self.vectorizer = TfidfVectorizer()
-        self.svd = TruncatedSVD(n_components=n_components)
-        self.embeddings = None
-        self.fitted = False
-
-    def add(self, text: str):
-        self.memory_texts.append(text)
-        self._update_embeddings()
-
-    def _update_embeddings(self):
-        if len(self.memory_texts) == 0:
-            self.embeddings = None
-            self.fitted = False
-            return
-
-        X = self.vectorizer.fit_transform(self.memory_texts)
-        n_comp = min(self.svd.n_components, X.shape[1], len(self.memory_texts)-1)
-        if n_comp <= 0:
-            self.embeddings = X.toarray()
-            self.fitted = True
-            return
-
-        self.svd = TruncatedSVD(n_components=n_comp)
-        self.embeddings = self.svd.fit_transform(X)
-        self.fitted = True
-
-    def retrieve(self, query: str, top_k=3):
-        if not self.fitted or self.embeddings is None or len(self.memory_texts) == 0:
-            return []
-
-        Xq = self.vectorizer.transform([query])
-        if self.svd.n_components > Xq.shape[1] or self.svd.n_components > len(self.memory_texts) - 1:
-            q_emb = Xq.toarray()
-        else:
-            q_emb = self.svd.transform(Xq)
-
-        sims = cosine_similarity(q_emb, self.embeddings)[0]
-        top_indices = sims.argsort()[::-1][:top_k]
-
-        return [self.memory_texts[i] for i in top_indices]
-
-    def process_input(self, new_text: str, top_k=3):
-        """자동으로 기억 저장하고, 유사한 기억 찾아서 합친 프롬프트 생성"""
-        related_memories = self.retrieve(new_text, top_k=top_k)
-        self.add(new_text)
-        return self.merge_prompt(new_text, related_memories)
-
-    def merge_prompt(self, prompt: str, memories: list):
-        context = "\n".join(memories)
-        full_prompt = ""
-        if context:
-            full_prompt += f"기억:\n{context}\n\n"
-        full_prompt += f"현재 질문:\n{prompt}\n\n응답:"
-        return full_prompt
-
-memory = SimilarityMemory()
-
-with open("base_texts.txt", "r", encoding="utf-8") as f:
-    for line in f:
-        line = line.strip()
-        if line:
-            memory.add(line)
-
 def mismatch_tone(input_text, output_text):  
     if "ㅋㅋ" in input_text and not re.search(r'ㅋㅋ|ㅎ|재밌|놀|만나|맛집|여행', output_text):  
         return True  
@@ -295,39 +224,30 @@ def is_valid_response(response):
 
 
 def respond(input_text):
-    memory.process_input(input_text)
-
+    # 이름 관련 질문에 딱 반응하는 부분 유지
     if "이름" in input_text:
         response = "제 이름은 Flexi입니다."
-        memory.process_input(response)
         return response
 
     if "누구" in input_text:
         response = "저는 Flexi라고 해요."
-        memory.process_input(response)
         return response
 
-    related_memories = memory.retrieve(input_text, top_k=3)
-    merged_prompt = memory.merge_prompt(input_text, related_memories)
+    # 메모리 관련 부분 싹 제거하고, 단순 프롬프트 생성
+    full_prompt = f"현재 질문:\n{input_text}\n\n응답:"
 
     for _ in range(3):  # 최대 3번 재시도
-        full_response = generate_text_sample(model, merged_prompt)
+        full_response = generate_text_sample(model, full_prompt)
 
-        # 여기서 '응답:' 뒤의 텍스트만 뽑기
         if "응답:" in full_response:
             response = full_response.split("응답:")[-1].strip()
         else:
             response = full_response.strip()
 
         if is_valid_response(response) and not mismatch_tone(input_text, response):
-            memory.process_input(response)
             return response
 
-    # 3번 모두 실패 시 fallback
-    fallback_response = "죄송해요, 제대로 답변을 못했어요."
-    memory.process_input(fallback_response)
-    return fallback_response
-
+    return "죄송해요, 제대로 답변을 못했어요."
 
 @app.get("/generate", response_class=PlainTextResponse)
 async def generate(request: Request):