Update api.py

api.py

@@ -140,19 +140,28 @@ _ = model(dummy_input)  # 모델이 빌드됨
 model.load_weights("InteractGPT.weights.h5")      
 print("모델 가중치 로드 완료!")      
 
+import re
+import math
+import numpy as np
+import requests
+import tensorflow as tf
+from sklearn.feature_extraction.text import TfidfVectorizer
+from sklearn.metrics.pairwise import cosine_similarity
+from fastapi import Request
+from fastapi.responses import PlainTextResponse
+
+# 1. Top-KP 기반 생성기
 def generate_text_topkp(model, prompt, max_len=100, max_gen=98,  
-                        temperature=0.50, min_len=20,  
+                        temperature=0.90, min_len=20,  
                         repetition_penalty=1.2, top_p=0.90, top_k=50):    
     def top_kp_filtering(logits, top_k, top_p):  
         probs = np.exp(logits - np.max(logits))  
         probs /= probs.sum()  
         sorted_idx = np.argsort(-probs)  
         sorted_probs = probs[sorted_idx]  
-        # Top-K 필터링  
         if top_k > 0:  
             sorted_idx = sorted_idx[:top_k]  
             sorted_probs = sorted_probs[:top_k]  
-        # Top-P 필터링  
         cum_probs = np.cumsum(sorted_probs)  
         cutoff = np.searchsorted(cum_probs, top_p) + 1  
         final_idx = sorted_idx[:cutoff]  
@@ -169,17 +178,13 @@ def generate_text_topkp(model, prompt, max_len=100, max_gen=98,
         input_tensor = tf.convert_to_tensor([input_padded])  
         logits = model(input_tensor, training=False)  
         next_logits = logits[0, len(generated) - 1].numpy()  
-        # 반복 억제  
         for t in set(generated):  
             count = generated.count(t)  
             next_logits[t] /= (repetition_penalty ** count)  
-        # 조기 종료 방지  
         if len(generated) < min_len:  
             next_logits[end_id] -= 5.0  
         next_logits[pad_id] -= 10.0  
-        # 온도 적용  
         next_logits = next_logits / temperature  
-        # Top-KP Sampling 적용  
         final_idx, final_probs = top_kp_filtering(next_logits, top_k=top_k, top_p=top_p)  
         sampled = np.random.choice(final_idx, p=final_probs)  
         generated.append(int(sampled))  
@@ -189,7 +194,37 @@ def generate_text_topkp(model, prompt, max_len=100, max_gen=98,
         decoded = decoded.strip()  
         if len(generated) >= min_len and (sampled == end_id or decoded.endswith(('.', '!', '?'))):  
             return decoded
+    return sp.decode(generated)
+
+# Greedy 버전 생성기
+def generate_text_greedy(model, prompt, max_len=100, max_gen=98):  
+    model_input = text_to_ids(f"<start> {prompt} <sep>")  
+    model_input = model_input[:max_len]  
+    generated = list(model_input)  
+    for _ in range(max_gen):  
+        pad_len = max(0, max_len - len(generated))  
+        input_padded = np.pad(generated, (0, pad_len), constant_values=pad_id)  
+        input_tensor = tf.convert_to_tensor([input_padded])  
+        logits = model(input_tensor, training=False)  
+        next_logits = logits[0, len(generated) - 1].numpy()  
+        next_logits[pad_id] -= 10.0  
+        next_token = np.argmax(next_logits)  
+        generated.append(int(next_token))  
+        decoded = sp.decode(generated)  
+        for t in ["<start>", "<sep>", "<end>"]:  
+            decoded = decoded.replace(t, "")  
+        decoded = decoded.strip()  
+        if next_token == end_id or decoded.endswith(('.', '!', '?')):  
+            return decoded
+    return sp.decode(generated)
+
+# 톤 불일치 체크
+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  
+    return False
 
+# 유효한 응답인지 검사
 def is_valid_response(response):
     if len(response.strip()) < 2:
         return False
@@ -203,6 +238,7 @@ def is_valid_response(response):
         return False
     return True
 
+# 위키 요약 관련
 def extract_main_query(text):
     sentences = re.split(r'[.?!]\s*', text)
     sentences = [s.strip() for s in sentences if s.strip()]
@@ -215,13 +251,6 @@ def extract_main_query(text):
         last = re.sub(rf'\b(\w+){p}\b', r'\1', last)
     return last.strip()
 
-import re
-import requests
-import numpy as np
-from sklearn.feature_extraction.text import TfidfVectorizer
-from sklearn.metrics.pairwise import cosine_similarity
-
-# 2. 위키백과 요약 가져오기
 def get_wikipedia_summary(query):
     cleaned_query = extract_main_query(query)
     url = f"https://ko.wikipedia.org/api/rest_v1/page/summary/{cleaned_query}"
@@ -231,24 +260,20 @@ def get_wikipedia_summary(query):
     else:
         return "위키백과에서 정보를 가져올 수 없습니다."
 
-# 3. TextRank 요약기
 def textrank_summarize(text, top_n=3):
     sentences = re.split(r'(?<=[.!?])\s+', text.strip())
     sentences = [s.strip() for s in sentences if len(s.strip()) > 10]
-
     if len(sentences) <= top_n:
-        return text  # 문장이 너무 적으면 원문 반환
-
+        return text
     vectorizer = TfidfVectorizer()
     tfidf_matrix = vectorizer.fit_transform(sentences)
     sim_matrix = cosine_similarity(tfidf_matrix)
     np.fill_diagonal(sim_matrix, 0)
-
     def pagerank(matrix, damping=0.85, max_iter=100, tol=1e-4):
         N = matrix.shape[0]
         ranks = np.ones(N) / N
         row_sums = np.sum(matrix, axis=1)
-        row_sums[row_sums == 0] = 1  # NaN 방지용 처리
+        row_sums[row_sums == 0] = 1
         for _ in range(max_iter):
             prev_ranks = ranks.copy()
             for i in range(N):
@@ -257,25 +282,22 @@ def textrank_summarize(text, top_n=3):
             if np.linalg.norm(ranks - prev_ranks) < tol:
                 break
         return ranks
-
     scores = pagerank(sim_matrix)
     ranked_idx = np.argsort(scores)[::-1]
     selected_idx = sorted(ranked_idx[:top_n])
     summary = ' '.join([sentences[i] for i in selected_idx])
-
     return summary
 
-# 4. 전체 파이프라인
 def summarize_from_wikipedia(query, top_n=3):
     raw_summary = get_wikipedia_summary(query)
     return textrank_summarize(raw_summary, top_n=top_n)
 
+# 의도 분류기
 def simple_intent_classifier(text):
     text = text.lower()
     greet_keywords = ["안녕", "반가워", "이름", "누구", "소개", "어디서 왔", "정체", "몇 살", "너 뭐야"]
     info_keywords = ["설명", "정보", "무엇", "뭐야", "어디", "누구", "왜", "어떻게", "종류", "개념"]
     math_keywords = ["더하기", "빼기", "곱하기", "나누기", "루트", "제곱", "+", "-", "*", "/", "=", "^", "√", "계산", "몇이야", "얼마야"]
-
     if any(kw in text for kw in greet_keywords):
         return "인사"
     elif any(kw in text for kw in info_keywords):
@@ -294,7 +316,7 @@ def parse_math_question(text):
     except:
         return "계산할 수 없는 수식이에요. 다시 한번 확인해 주세요!"
 
-# 전체 응답 함수
+# 최종 응답 함수
 def respond(input_text):
     intent = simple_intent_classifier(input_text)
 
@@ -309,7 +331,7 @@ def respond(input_text):
 
     if intent == "인사":
         return "반가워요! 무엇을 도와드릴까요?"
-        
+
     if intent == "정보질문":
         keyword = re.sub(r"(에 대해|에 대한|에 대해서)?\s*(설명해줘|알려줘|뭐야|개념|정의|정보)?", "", input_text).strip()
         if not keyword:
@@ -317,16 +339,11 @@ def respond(input_text):
         summary = summarize_from_wikipedia(keyword)
         return f"{summary}\n다른 궁금한 점 있으신가요?"
 
-    return generate_text_topkp(model, input_text)
-
-async def async_generator_wrapper(prompt: str):
-    gen = generate_text_topkp(model, prompt)
-    for text_piece in gen:
-        yield text_piece
-        await asyncio.sleep(0.1)
-
-
-from fastapi.responses import PlainTextResponse
+    # 일상 대화: 샘플링 + fallback
+    response = generate_text_topkp(model, input_text)
+    if not is_valid_response(response) or mismatch_tone(input_text, response):
+        response = generate_text_greedy(model, input_text)
+    return response
 
 @app.get("/generate", response_class=PlainTextResponse)
 async def generate(request: Request):