Update app.py

app.py

@@ -45,40 +45,71 @@ LANG_MEMBER_MAP = {
 }
 
 async def band_consulting(user_input, selected_lang, g_inst, b_inst, d_inst, chords):
-    # ... (생략: LLM 답변 생성 로직 동일)
-    
-    # 1. 파싱 로직 정교화
-    tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
-    music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
-    trans_match = re.search(r'\[TRANSLATION\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
-    
-    tab_display = tab_match.group(1).strip() if tab_match else "No Tab Data"
-    translation = trans_match.group(1).strip() if trans_match else ""
-    
-    # ★ 핵심: 맴버가 읽을 텍스트(TTS)는 번역문을 제외한 clean_text만 사용
-    clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL | re.IGNORECASE)
-    clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text, flags=re.IGNORECASE)
-    clean_text = re.sub(r'\[TRANSLATION\].*?(\[|$)', '', clean_text, flags=re.DOTALL | re.IGNORECASE).strip()
-    
-    # TTS 생성 (번역문 없이 네이티브 설명만 읽음)
-    m_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
-    # 특수문자 제거 후 순수 텍스트만 전달
-    tts_text = re.sub(r'[\*\#\-\_\~\|]', '', clean_text)
-    communicate = edge_tts.Communicate(tts_text, m_info["voice"])
-    await communicate.save(voice_path)
+    try:
+        req_id = str(uuid.uuid4())[:8]
+        voice_path = f"/tmp/v_{req_id}.mp3"
+        music_path = f"/tmp/m_{req_id}.wav"
+        
+        m_info = LANG_MEMBER_MAP.get(selected_lang, LANG_MEMBER_MAP["Korean"])
+        
+        # 프롬프트 강화: 5-7줄 설명 및 영어 번역 섹션 강제
+        system_prompt = f"""You are {m_info['name']}. Respond ONLY in {selected_lang}.
+        Provide 5-7 lines of professional music advice. 
+        [TAB] Section: Detailed chords/tabs.
+        [TRANSLATION] English translation of your advice.
+        [MUSIC] English prompt: Guitar:{g_inst}, Bass:{b_inst}, Drums:{d_inst}, Chords:{chords}"""
+        
+        ai_text_raw = ""
+        groq_client = ModelManager.get_groq()
+        if groq_client:
+            try:
+                res = groq_client.chat.completions.create(
+                    messages=[{"role": "system", "content": system_prompt}, {"role": "user", "content": user_input}],
+                    model="llama-3.3-70b-versatile"
+                )
+                ai_text_raw = res.choices[0].message.content
+            except: pass
 
-    # MusicGen (12초)
-    music_gen = ModelManager.get_music()
-    music_p = music_match.group(1).strip() if music_match else "rock"
-    music_output = music_gen(music_p, forward_params={"max_new_tokens": 512})
-    audio_data = np.squeeze(music_output["audio"])
-    audio_int16 = (audio_data * 32767).astype(np.int16)
-    scipy.io.wavfile.write(music_path, music_output["sampling_rate"], audio_int16)
+        if not ai_text_raw:
+            qwen = ModelManager.get_qwen()
+            input_t = f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{user_input}<|im_end|>\nassistant\n"
+            out = qwen(input_t, max_new_tokens=1024)
+            ai_text_raw = out[0]['generated_text'].split("assistant\n")[-1]
 
-    # 리턴 시 번역 포함 (clean_text, voice_path, music_path, tab_display, translation)
-    return clean_text, voice_path, music_path, tab_display, translation
+        # 파싱 로직
+        tab_match = re.search(r'\[TAB\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
+        music_match = re.search(r'\[MUSIC:(.*?)\]', ai_text_raw, re.IGNORECASE)
+        trans_match = re.search(r'\[TRANSLATION\](.*?)(\[|$)', ai_text_raw, re.DOTALL | re.IGNORECASE)
+        
+        tab_display = tab_match.group(1).strip() if tab_match else "No Tab Data"
+        translation = trans_match.group(1).strip() if trans_match else ""
+        
+        # TTS용 클린 텍스트 (번역문/음악프롬프트/악보 제외)
+        clean_text = re.sub(r'\[TAB\].*?(\[|$)', '', ai_text_raw, flags=re.DOTALL | re.IGNORECASE)
+        clean_text = re.sub(r'\[MUSIC:.*?\]', '', clean_text, flags=re.IGNORECASE)
+        clean_text = re.sub(r'\[TRANSLATION\].*?(\[|$)', '', clean_text, flags=re.DOTALL | re.IGNORECASE).strip()
+        
+        # TTS 생성
+        tts_text = re.sub(r'[\*\#\-\_\~\|]', '', clean_text)
+        communicate = edge_tts.Communicate(tts_text, m_info["voice"])
+        await communicate.save(voice_path)
+
+        # MusicGen (12초)
+        music_gen = ModelManager.get_music()
+        music_p = music_match.group(1).strip() if music_match else "rock"
+        music_output = music_gen(music_p, forward_params={"max_new_tokens": 512})
+        audio_data = np.squeeze(music_output["audio"])
+        audio_int16 = (audio_data * 32767).astype(np.int16)
+        scipy.io.wavfile.write(music_path, music_output["sampling_rate"], audio_int16)
+
+        return clean_text, voice_path, music_path, tab_display, translation
+
+    except Exception as e:
+        print(f"Error: {str(e)}")
+        return f"Error: {str(e)}", None, None, "No Tab", "Error occurred"
 
 with gr.Blocks() as demo:
+    # 인풋 6개 (script.js의 predict 배열과 일치)
     inputs = [gr.Textbox(visible=False) for _ in range(6)]
     outputs = [gr.Textbox(visible=False), gr.Audio(visible=False), gr.Audio(visible=False), gr.Textbox(visible=False), gr.Textbox(visible=False)]
     btn = gr.Button("API", visible=False)