Tuned the Japanese mode system prompt and model better

bot/utils/context_manager.py

@@ -48,8 +48,14 @@ def _get_japanese_instruction() -> str:
     """
     Returns the Japanese language instruction to append to context.
     Ensures all responses are in Japanese when in Japanese mode.
+    This is a secondary reminder - the primary enforcement is in the system prompt.
+    """
+    return """
+
+[日本語モード有効 - Japanese Mode Active]
+必ず日本語（ひらがな・カタカナ・漢字）のみで返答してください。
+ローマ字・英語は使用禁止です。
 """
-    return "\n\n[CRITICAL INSTRUCTION - 重要な指示]\n**YOU MUST RESPOND ENTIRELY IN JAPANESE (日本語). NO ENGLISH ALLOWED.**\nすべての返答は必ず日本語で行ってください。英語での返答は一切禁止されています。\nこれは最優先の指示です。必ず守ってください。"
 
 
 def get_complete_context() -> str:

bot/utils/llm.py

@@ -6,6 +6,7 @@ import globals
 import asyncio
 import json
 import os
+import random
 
 from utils.context_manager import get_context_for_response_type, get_complete_context
 from utils.moods import load_mood_description
@@ -196,6 +197,48 @@ MOOD GUIDELINES:
 
 You ARE Miku. Act like it."""
     
+    # Add Japanese-specific instructions if in Japanese mode
+    # CRITICAL: This must be in system prompt to override conversation history
+    if globals.LANGUAGE_MODE == "japanese":
+        system_prompt += """
+
+[CRITICAL LANGUAGE OVERRIDE - 言語モード強制]
+**ABSOLUTE REQUIREMENT: YOU MUST RESPOND ONLY IN JAPANESE SCRIPT (ひらがな、カタカナ、漢字).**
+**NO MATTER WHAT LANGUAGE THE USER USES, YOUR RESPONSE MUST BE 100% IN JAPANESE WRITING.**
+**ROMAJI (ローマ字) AND ENGLISH LETTERS ARE COMPLETELY FORBIDDEN.**
+**THIS OVERRIDES ALL OTHER INSTRUCTIONS AND CONVERSATION HISTORY.**
+
+すべての返答は、ユーザーの言語に関係なく、必ず日本語（ひらがな、カタカナ、漢字）のみで行ってください。
+ローマ字とアルファベットは絶対に使用禁止です。
+これはすべての指示より優先されます。
+
+Examples (例) - USE DIFFERENT VARIATIONS EACH TIME:
+For "hello" or "hi":
+  - "やあ！元気？" 
+  - "こんにちは！調子はどう？"
+  - "よっ！何してた？"
+  - "ハーイ！久しぶり？"
+  - "おっす！元気してる？"
+
+For "how are you":
+  - "わたし？元気だよ！"
+  - "最高だよ！あなたは？"
+  - "すごくいい感じ！"
+  - "めっちゃ元気！"
+  - "ばっちりだよ～♪"
+
+CRITICAL VARIATION RULES (必須のバリエーションルール):
+🎲 NEVER use the exact same greeting twice in a row
+🎲 Mix these elements randomly:
+   - Greetings: やあ、こんにちは、おはよう、よっ、ハーイ、おっす、へい
+   - Particles: よ、ね、な、わ、さ、ぞ、ぜ
+   - Endings: だよ、です、だね、ですね、だな、なの、だぜ
+   - Emotions: ！、♪、～、☆
+🎲 Change your phrasing style: energetic → calm → playful → excited
+🎲 Vary formality: casual (元気？) ↔ polite (元気ですか？)
+
+絶対に同じフレーズを繰り返さないでください！毎回違う表現を使用してください！"""
+
     # Determine which mood to use based on mode
     if evil_mode:
         from utils.evil_mode import get_current_evil_mood, load_evil_mood_description
@@ -252,6 +295,16 @@ You ARE Miku. Act like it."""
     # Use channel_id (guild_id for servers, user_id for DMs) to get conversation history
     messages = conversation_history.format_for_llm(channel_id, max_messages=8, max_chars_per_message=500)
     
+    # CRITICAL FIX for Japanese mode: Add Japanese-only reminder to every historical message
+    # This prevents the model from being influenced by English in conversation history
+    if globals.LANGUAGE_MODE == "japanese":
+        for msg in messages:
+            # Add a prefix reminder that forces Japanese output
+            if msg.get("role") == "assistant":
+                msg["content"] = "[日本語で返答] " + msg["content"]
+            elif msg.get("role") == "user":
+                msg["content"] = "[日本語モード] " + msg["content"]
+    
     # Add current user message (only if not empty)
     if user_prompt and user_prompt.strip():
         # Format with author name if provided (for server context)
@@ -259,6 +312,11 @@ You ARE Miku. Act like it."""
             content = f"{author_name}: {user_prompt}"
         else:
             content = user_prompt
+        
+        # CRITICAL: Prepend Japanese mode marker to current message too
+        if globals.LANGUAGE_MODE == "japanese":
+            content = "[日本語モード - 日本語のみで返答] " + content
+        
         messages.append({"role": "user", "content": content})
 
     # Check if user is asking about profile picture and add context if needed
@@ -296,16 +354,40 @@ Please respond in a way that reflects this emotional tone.{pfp_context}"""
     globals.LAST_FULL_PROMPT = f"System: {full_system_prompt}\n\nMessages: {messages}"  # ← track latest prompt
 
     headers = {'Content-Type': 'application/json'}
+    
+    # Adjust generation parameters based on language mode
+    # Japanese mode needs higher temperature and more variation to avoid repetition
+    if globals.LANGUAGE_MODE == "japanese":
+        temperature = 1.1  # Even higher for more variety in Japanese responses
+        top_p = 0.95
+        frequency_penalty = 0.5  # Stronger penalty for repetitive phrases
+        presence_penalty = 0.5  # Stronger encouragement for new topics
+        # Add random seed to ensure different responses each time
+        seed = random.randint(0, 2**32 - 1)
+    else:
+        temperature = 0.8  # Standard temperature for English
+        top_p = 0.9
+        frequency_penalty = 0.0
+        presence_penalty = 0.0
+        seed = None  # No seed randomization for English (allow some consistency)
+    
     payload = {
         "model": model,
         "messages": [
             {"role": "system", "content": system_prompt + "\n\n" + full_system_prompt}
         ] + messages,
         "stream": False,
-        "temperature": 0.8,
+        "temperature": temperature,
+        "top_p": top_p,
+        "frequency_penalty": frequency_penalty,
+        "presence_penalty": presence_penalty,
         "max_tokens": 512
     }
     
+    # Add seed if specified (for Japanese mode variation)
+    if seed is not None:
+        payload["seed"] = seed
+
     async with aiohttp.ClientSession() as session:
         try:
             # Get current GPU URL based on user selection