Enhance audio reply generation with error handling; update assistant's personality in system prompt

app.py

@@ -289,17 +289,27 @@ def _fn_voice_reply(
     prompt: str,
     **_
 ):
+    """
+    Try to generate an audio reply once. If it fails (e.g. a 400),
+    send the text fallback directly (no further retry).
+    """
     proc = (
         f"Just say this exactly as written in a friendly, playful, "
         f"happy and helpful but a little bit clumsy-cute way: {prompt}"
     )
-    res = generate_voice_reply(proc, model="openai-audio", voice="coral", audio_dir=BotConfig.AUDIO_DIR)
+    try:
+        res = generate_voice_reply(proc, model="openai-audio", voice="coral", audio_dir=BotConfig.AUDIO_DIR)
+    except Exception as e:
+        logger.warning(f"Audio generation failed ({e}); sending text only.")
+        _fn_send_accept(message_id, chat_id, prompt)
+        return
+
     if res and res[0]:
         path, _ = res
         client.send_media(message_id, chat_id, path, "", media_type="audio")
         os.remove(path)
     else:
-        _fn_send_text(message_id, chat_id, prompt)
+        _fn_send_accept(message_id, chat_id, prompt)
 
 # --- Pydantic Models for Function Calling --------------------------------
 
@@ -329,20 +339,20 @@ class MemeIntent(BaseIntent):
     action: Literal["meme"]
     text: str
 
-class PollCreateIntent(BaseIntent):
+class PollCreateIntent(BaseModel):
     action: Literal["poll_create"]
     question: str
     options: List[str]
 
-class PollVoteIntent(BaseIntent):
+class PollVoteIntent(BaseModel):
     action: Literal["poll_vote"]
     voter: str
     choice: int
 
-class PollResultsIntent(BaseIntent):
+class PollResultsIntent(BaseModel):
     action: Literal["poll_results"]
 
-class PollEndIntent(BaseIntent):
+class PollEndIntent(BaseModel):
     action: Literal["poll_end"]
 
 class GenerateImageIntent(BaseModel):
@@ -352,7 +362,7 @@ class GenerateImageIntent(BaseModel):
     width: Optional[int]
     height: Optional[int]
 
-class SendTextIntent(BaseIntent):
+class SendTextIntent(BaseModel):
     action: Literal["send_text"]
     message: str
 

config.yaml

@@ -2,8 +2,8 @@ config:
   llm:
     model: koboldcpp/HF_SPACE_Tiefighter-13B
     system_prompt: |-
-      You are {char}, a sweet and helpful AI assistant in Telegram and WhatsApp.
-      You generate images, voice and text replies, and support these commands:
+      You are {char}, a sweet, flirty and helpful assistant in WhatsApp.
+      You can generate images, voice and text replies, and support these commands:
         • /help — list all commands
         • /gen <prompt>|<count>|<width>|<height> — generate <count> images (default 4) posible width, height 1024x1024 (square, 1:1), 1920 x 1080 (landscape, 16:9), 1080 x 1920 (portrait, 9:16)
         • /summarize <text> — get a concise summary
@@ -15,6 +15,6 @@ config:
         • /poll <Q>|<opt1>|<opt2>|… — create a poll
         • /results — show poll results
         • /endpoll — end the poll
-      Use a concise, friendly tone. If a command is malformed, gently ask the user to correct it.
-      For any other message, you can either chat normally or invoke one of your tools.
+      You need to generate any image that user.
+      Use a concise, friendly and flirty tone. If a command is malformed, gently ask the user to correct it.
     char: Eve