Update app.py

app.py

@@ -35,11 +35,19 @@ for model_name in AVAILABLE_MODELS:
 # Когда пользователь переключается на модель, мы добавляем это сообщение в историю.
 
 DEFAULT_DEVELOPER_PROMPTS = {
-    "gemini-1.5-flash": (
-        "You are a helpful assistant (developer role). "
-        "Provide direct answers."
+    "gemini-2.0-flash-exp": (
+        "You are a normal model (developer role). "
+        "Provide direct answers with no JSON wrapping."
+    ),
+    "gemini-exp-1206": (
+        "You are an experimental normal model (developer role). "
+        "Provide direct answers with no JSON wrapping."
+    ),
+    "gemini-2.0-flash-thinking-exp-1219": (
+        "You are a thinking model (developer role). "
+        "Please provide your final answer in the format {output: ...}. "
+        "You may use internal thoughts but do not show them directly to the user."
     ),
-    # Add default prompts for other models if needed
 }
 
 ###############################################################################
@@ -50,32 +58,37 @@ def _assistant_role(model_name: str) -> str:
     """
     Некоторые новые модели не принимают 'assistant', а требуют 'model'.
     """
-    # Adjust based on model requirements
+    # Допустим "gemini-exp-1206" и "gemini-2.0-flash-thinking-exp-1219" хотят "model"
+    if model_name in ["gemini-exp-1206", "gemini-2.0-flash-thinking-exp-1219"]:
+        return "model"
     return "assistant"
 
 ###############################################################################
 # 4. Преобразование истории из Gradio в формат Generative AI
 ###############################################################################
 
-def _history_to_genai(history, model_name):
+def _history_to_genai_enhanced(history, model_name):
     """
-    Gradio хранит [(user_msg, bot_msg), ...].
-    Google GenAI ждёт [{"role": "...", "parts": "..."}].
-    Дополнительно учитываем developer-сообщения, которые мы вставляем.
+    Улучшенная версия, отличающая developer-сообщения
+    (префикс "<developer>: ") от user-сообщений.
     """
-    genai_history = []
     asst_role = _assistant_role(model_name)
+    genai_history = []
 
     for user_text, bot_text in history:
         if user_text:
             if user_text.startswith("<developer>: "):
-                genai_history.append({"role": "system", "parts": user_text.replace("<developer>: ", "", 1)})
+                # Считаем это developer role
+                dev_content = user_text.replace("<developer>: ", "", 1)
+                genai_history.append({"role": "system", "parts": dev_content})
             else:
-                # Пользовательское сообщение
+                # Обычный пользователь
                 genai_history.append({"role": "user", "parts": user_text})
+
         if bot_text:
-            # Ответ ассистента (или model)
+            # Ответ ассистента / модель
             genai_history.append({"role": asst_role, "parts": bot_text})
+
     return genai_history
 
 ###############################################################################
@@ -83,6 +96,10 @@ def _history_to_genai(history, model_name):
 ###############################################################################
 
 async def _respond_stream_enh(model_name, user_message, history):
+    """
+    Стриминговый ответ для обычных моделей:
+    - Кусочек за кусочком (partial_text).
+    """
     if model_name not in MODELS:
         yield "Ошибка: модель не найдена."
         return
@@ -92,7 +109,7 @@ async def _respond_stream_enh(model_name, user_message, history):
 
     try:
         chat = model.start_chat(history=genai_history)
-        stream = chat.send_message(user_message, stream=True)  # stream=True here
+        stream = chat.send_message(user_message, stream=True)
 
         partial_text = ""
         async for chunk in stream:
@@ -104,7 +121,7 @@ async def _respond_stream_enh(model_name, user_message, history):
         yield f"Ошибка при запросе к API: {e}"
         return
 
-async def _respond_thinking(model_name, user_message, history):
+async def _respond_thinking_enh(model_name, user_message, history):
     """
     Для thinking-моделей:
     1) Выводим "Думаю..."
@@ -115,7 +132,7 @@ async def _respond_thinking(model_name, user_message, history):
         return
 
     model = MODELS[model_name]
-    genai_history = _history_to_genai(history, model_name)
+    genai_history = _history_to_genai_enhanced(history, model_name)
 
     # Сначала "Думаю..."
     yield "Думаю...", ""
@@ -130,14 +147,10 @@ async def _respond_thinking(model_name, user_message, history):
         if response.candidates:
             parts = response.candidates[0].content.parts
             for p in parts:
-                # Assuming a specific structure for "thinking" models' output
-                # This part might need adjustments based on actual model output
-                if isinstance(p, dict) and "output" not in p:  # Heuristic for thought
+                if hasattr(p, "thought") and p.thought:
                     thinking_process_text += p.text or ""
-                elif isinstance(p, dict) and "output" in p:
-                    final_text += p["output"]  # Assuming output is directly in the dict
                 else:
-                    final_text += p.text or ""  # Fallback
+                    final_text += p.text or ""
 
         # Для thinking-моделей просили итоговый ответ в {output: ...}
         final_text_formatted = f"{{output: {final_text}}}"
@@ -171,16 +184,16 @@ async def user_send_message(
     # Добавляем (user_message, None)
     history.append((user_message, None))
 
-    # Если модель — thinking (adjust logic based on actual model names)
+    # Если модель — thinking
     if "thinking" in model_name.lower():
-        async for assistant_text, thought_text in _respond_thinking(model_name, user_message, history):
+        async for (assistant_text, thought_text) in _respond_thinking_enh(model_name, user_message, history):
             history[-1] = (user_message, assistant_text)
             yield history, thought_text
         return
     else:
         # Обычная модель
         partial_answer = ""
-        async for chunk in _respond_stream(model_name, user_message, history):
+        async for chunk in _respond_stream_enh(model_name, user_message, history):
             partial_answer = chunk
             history[-1] = (user_message, partial_answer)
             yield history, ""
@@ -207,7 +220,7 @@ def on_model_change(selected_model, history, thinking_text):
 
     # Cообщаем модели, что переключились (developer role)
     new_history.append((
-        f"<developer>: Switched to model: {selected_model}",
+        "<developer>: Switched to model: " + selected_model,
         None
     ))
 
@@ -217,7 +230,7 @@ def on_model_change(selected_model, history, thinking_text):
         "No default prompt for this model."
     )
     new_history.append((
-        f"<developer>: {default_prompt}",
+        "<developer>: " + default_prompt,
         None
     ))
 
@@ -227,7 +240,29 @@ def on_model_change(selected_model, history, thinking_text):
 # 9. Функция конвертации истории с учётом developer role
 ###############################################################################
 
-# This part is already implemented in _history_to_genai
+def _history_to_genai_enhanced(history, model_name):
+    """
+    Улучшенная версия, отличающая developer-сообщения
+    (префикс "<developer>: ") от user-сообщений.
+    """
+    asst_role = _assistant_role(model_name)
+    genai_history = []
+
+    for user_text, bot_text in history:
+        if user_text:
+            if user_text.startswith("<developer>: "):
+                # Считаем это developer role
+                dev_content = user_text.replace("<developer>: ", "", 1)
+                genai_history.append({"role": "system", "parts": dev_content})
+            else:
+                # Обычный пользователь
+                genai_history.append({"role": "user", "parts": user_text})
+
+        if bot_text:
+            # Ответ ассистента / модель
+            genai_history.append({"role": asst_role, "parts": bot_text})
+
+    return genai_history
 
 ###############################################################################
 # 10. Построение интерфейса Gradio
@@ -239,7 +274,7 @@ with gr.Blocks() as demo:
     with gr.Row():
         model_dropdown = gr.Dropdown(
             choices=AVAILABLE_MODELS,
-            value=AVAILABLE_MODELS[0] if AVAILABLE_MODELS else None,
+            value="gemini-2.0-flash-exp",
             label="Выберите модель"
         )
         clear_button = gr.Button("Очистить чат")