Update api.py

api.py

@@ -10,10 +10,52 @@ import re
 import math
 from sklearn.feature_extraction.text import TfidfVectorizer
 from sklearn.metrics.pairwise import cosine_similarity
+from sentence_transformers import SentenceTransformer
+import faiss
 
-app = FastAPI()    
+app = FastAPI()  
 
+class SimilarityMemory:
+    def __init__(self, embed_model='all-MiniLM-L6-v2'):
+        self.memory_texts = []
+        self.model = SentenceTransformer(embed_model)
+        self.index = None
+        self.vectors = []
 
+    def add(self, text: str):
+        vec = self.model.encode([text])[0]
+        self.memory_texts.append(text)
+        self.vectors.append(vec)
+        self._update_index()
+
+    def _update_index(self):
+        if not self.vectors:
+            return
+        vecs = np.vstack(self.vectors).astype('float32')
+        self.index = faiss.IndexFlatL2(vecs.shape[1])
+        self.index.add(vecs)
+
+    def retrieve(self, query: str, top_k=3):
+        if not self.vectors:
+            return []
+        q_vec = self.model.encode([query])[0].astype('float32').reshape(1, -1)
+        D, I = self.index.search(q_vec, top_k)
+        return [self.memory_texts[i] for i in I[0]]
+
+memory = SimilarityMemory()
+
+# 예시 기억 추가 (이건 운영 중에 실시간으로 갱신 가능)
+memory.add("어제는 기분이 별로였어")
+memory.add("영화 보러 갔다가 친구랑 싸웠어")
+memory.add("카페에서 공부했는데 집중이 잘 됐어")
+
+
+def merge_prompt_with_memory(prompt: str, memories: list):
+    context = "\n".join(f"과거: {mem}" for mem in memories)
+    return f"{context}\n현재: {prompt}"
+
+
+        
 from fastapi.middleware.cors import CORSMiddleware
 
 origins = [
@@ -330,7 +372,6 @@ def parse_math_question(text):
     except:
         return "계산할 수 없는 수식이에요. 다시 한번 확인해 주세요!"
 
-# 최종 응답 함수
 def respond(input_text):
     intent = simple_intent_classifier(input_text)
 
@@ -350,12 +391,16 @@ def respond(input_text):
         summary = summarize_from_wikipedia(keyword)
         return f"{summary}\n다른 궁금한 점 있으신가요?"
 
-    # 일상 대화: 샘플링 + fallback
-    response = generate_text_sample(model, input_text)
+    # ✅ 기억 기반 병합 추가
+    related_memories = memory.retrieve(input_text, top_k=3)
+    merged_prompt = merge_prompt_with_memory(input_text, related_memories)
+
+    response = generate_text_sample(model, merged_prompt)
     if not is_valid_response(response) or mismatch_tone(input_text, response):
-        response = generate_text_sample(model, input_text)
+        response = generate_text_sample(model, merged_prompt)
     return response
 
+
 @app.get("/generate", response_class=PlainTextResponse)
 async def generate(request: Request):
     prompt = request.query_params.get("prompt", "안녕하세요")