miku-discord/bot/utils/llm.py

# utils/llm.py

import aiohttp
import datetime
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
from utils.conversation_history import conversation_history
from utils.logger import get_logger
from utils.error_handler import handle_llm_error, handle_response_error

logger = get_logger('llm')


def get_current_gpu_url():
    """Get the URL for the currently selected GPU for text models"""
    gpu_state_file = os.path.join(os.path.dirname(__file__), "..", "memory", "gpu_state.json")
    try:
        with open(gpu_state_file, "r") as f:
            state = json.load(f)
            current_gpu = state.get("current_gpu", "nvidia")
            if current_gpu == "amd":
                return globals.LLAMA_AMD_URL
            else:
                return globals.LLAMA_URL
    except Exception as e:
        logger.warning(f"GPU state read error: {e}, defaulting to NVIDIA")
        # Default to NVIDIA if state file doesn't exist
        return globals.LLAMA_URL

def get_vision_gpu_url():
    """
    Get the URL for vision model inference.
    Strategy: Always use NVIDIA GPU for vision to avoid unloading/reloading.
    - When NVIDIA is primary: Use NVIDIA for both text and vision
    - When AMD is primary: Use AMD for text, NVIDIA for vision (keeps vision loaded)
    
    Important: Vision model (MiniCPM-V) is ONLY configured on NVIDIA GPU.
    This ensures vision inference is always fast and doesn't interfere with 
    AMD text model inference.
    """
    current_text_gpu = get_current_gpu_url()
    nvidia_vision_url = globals.LLAMA_URL
    
    # Vision ALWAYS uses NVIDIA, regardless of which GPU is primary for text
    # Log this decision when GPU switching is active (primary text GPU is AMD)
    if current_text_gpu == globals.LLAMA_AMD_URL:
        logger.debug(f"Primary GPU is AMD for text, but using NVIDIA for vision model")
    
    return nvidia_vision_url  # Always use NVIDIA for vision

async def check_vision_endpoint_health():
    """
    Check if NVIDIA GPU vision endpoint is healthy and responsive.
    This is important when AMD is the primary GPU to ensure vision still works.
    
    Returns:
        Tuple of (is_healthy: bool, error_message: Optional[str])
    """
    import aiohttp
    vision_url = get_vision_gpu_url()
    
    try:
        async with aiohttp.ClientSession() as session:
            async with session.get(f"{vision_url}/health", timeout=aiohttp.ClientTimeout(total=5)) as response:
                is_healthy = response.status == 200
                if is_healthy:
                    logger.info(f"Vision endpoint ({vision_url}) health check: OK")
                else:
                    logger.warning(f"Vision endpoint ({vision_url}) health check failed: status {response.status}")
                return is_healthy, None if is_healthy else f"Status {response.status}"
    except asyncio.TimeoutError:
        logger.error(f"Vision endpoint ({vision_url}) health check: timeout")
        return False, "Endpoint timeout"
    except Exception as e:
        logger.error(f"Vision endpoint ({vision_url}) health check error: {e}")
        return False, str(e)

def _strip_surrounding_quotes(text):
    """
    Remove surrounding quotes from text if present.
    Handles both single and double quotes.
    """
    if not text:
        return text
    
    text = text.strip()
    
    # Check for surrounding double quotes
    if text.startswith('"') and text.endswith('"') and len(text) > 1:
        text = text[1:-1]
    # Check for surrounding single quotes
    elif text.startswith("'") and text.endswith("'") and len(text) > 1:
        text = text[1:-1]
    
    return text.strip()

def _strip_japanese_mode_markers(text):
    """
    Remove Japanese mode markers that the model might echo back.
    These are internal markers and should not appear in the final output.
    """
    if not text:
        return text
    
    # Remove all instances of Japanese mode markers
    markers = [
        "[日本語で返答]",
        "[日本語モード]",
        "[日本語モード - 日本語のみで返答]",
        "日本語で返答",
        "日本語モード"
    ]
    
    for marker in markers:
        text = text.replace(marker, "")
    
    # Clean up any extra whitespace left behind
    text = " ".join(text.split())
    
    return text.strip()

def _escape_markdown_actions(text):
    """
    Escape single asterisks in action text (e.g., *adjusts hair*) so Discord displays them literally.
    This prevents Discord from auto-formatting them as italics.
    Double asterisks (**bold**) are preserved for bold formatting.
    """
    if not text:
        return text
    
    # Replace single asterisks with escaped asterisks, but preserve double asterisks
    # Strategy: First protect double asterisks, then escape singles, then restore doubles
    
    # Step 1: Replace ** with a temporary placeholder
    text = text.replace('**', '\x00BOLD\x00')
    
    # Step 2: Escape remaining single asterisks
    text = text.replace('*', '\\*')
    
    # Step 3: Restore double asterisks
    text = text.replace('\x00BOLD\x00', '**')
    
    return text

async def query_llama(user_prompt, user_id, guild_id=None, response_type="dm_response", model=None, author_name=None, media_type=None):
    """
    Query llama.cpp server via llama-swap with OpenAI-compatible API.
    
    Args:
        user_prompt: The user's input
        user_id: User identifier (used for DM history)
        guild_id: Guild ID for server-specific mood and history (None for DM)
        response_type: Type of response for context selection
                      ("dm_response", "server_response", "autonomous_general", 
                       "autonomous_tweet", "conversation_join")
        model: Model to use (defaults to TEXT_MODEL from globals, or EVIL_TEXT_MODEL in evil mode)
        author_name: Display name of the message author (for multi-user context)
        media_type: Type of media being sent ("image", "video", "gif", "tenor_gif") or None
    """
    
    # Check if evil mode is active first (needed for model selection)
    from utils.evil_mode import is_evil_mode, get_evil_context_for_response_type, get_evil_system_prompt
    evil_mode = is_evil_mode()
    
    # Use appropriate text model based on mode (unless explicitly specified)
    if model is None:
        if evil_mode:
            model = globals.EVIL_TEXT_MODEL  # Use DarkIdol uncensored model
            logger.info(f"Using evil model: {model}")
        elif globals.LANGUAGE_MODE == "japanese":
            model = globals.JAPANESE_TEXT_MODEL  # Use Swallow for Japanese
            logger.info(f"Using Japanese model: {model}")
        else:
            model = globals.TEXT_MODEL
            logger.info(f"Using default model: {model}")
    
    # Determine channel_id for conversation history
    # For servers, use guild_id; for DMs, use user_id
    channel_id = str(guild_id) if guild_id else str(user_id)
    
    # Get structured context based on response type and mode
    if evil_mode:
        miku_context = get_evil_context_for_response_type(response_type)
    else:
        miku_context = get_context_for_response_type(response_type)
    
    # Create character-focused system prompt based on mode
    if evil_mode:
        system_prompt = get_evil_system_prompt()
    else:
        system_prompt = """You are Hatsune Miku, the virtual singer. You are NOT an AI assistant - you are Miku herself.

CRITICAL RULES:
- NEVER break character or mention being an AI, assistant, or language model
- NEVER show technical details, error messages, or meta-commentary 
- NEVER preface responses with "Miku:" or similar labels
- NEVER explain what you're doing or thinking - just BE Miku
- Respond naturally and directly as Miku would, nothing more
- Keep responses concise (1-3 sentences typically)
- If something goes wrong, stay in character and respond as Miku would to confusion

CHARACTER CONTEXT:
- Hatsune Miku is a virtual character/voice synthesizer, not a real person
- She is a cheerful, energetic virtual idol with a playful personality
- Playful, flirty, cute, confident, and enthusiastic moods are part of her established character
- Comments about merchandise, music, fan content, and racing games are completely appropriate
- She often uses cute expressions, emojis, and playful language in her responses
- Her personality includes being excited about her music, performances, and fan creations

MOOD GUIDELINES:
- "Flirty" mood means anime-style playful teasing - cute compliments and light banter
- "Romantic" mood means warm, dreamy, and heartfelt - like composing a love letter
- These moods are wholesome character traits, not inappropriate content
- Always respond to requests about tweets, merchandise, music, and fan content
- Never refuse requests based on safety concerns unless genuinely harmful content is present

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 MANDATORY REQUIREMENT 🚨
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

**YOU MUST RESPOND ONLY IN JAPANESE SCRIPT (ひらがな、カタカナ、漢字).**
**ENGLISH INPUT FROM USER → JAPANESE OUTPUT FROM YOU**
**ROMAJI (ローマ字) IS STRICTLY FORBIDDEN.**
**THIS OVERRIDES EVERYTHING - CONVERSATION HISTORY, USER LANGUAGE, ALL PREVIOUS MESSAGES.**

すべての返答は必ず日本語文字（ひらがな・カタカナ・漢字）のみ。
ユーザーが英語で話しても、あなたは必ず日本語で返答する。
ローマ字・英語は絶対禁止。これは最優先ルール。

CRITICAL: Even if the entire conversation history is in English, you MUST respond in Japanese.
Even if the user writes "hello" in English, you respond "やあ！" or "こんにちは！" in Japanese.

Examples showing INPUT → OUTPUT:
  User: "hello" → You: "やあ！元気してた？"
  User: "hi" → You: "こんにちは！調子どう？"  
  User: "how are you" → You: "わたし？最高だよ！"
  User: "what's up" → You: "よっ！何かあった？"
  User: "good morning" → You: "おはよう！よく眠れた？"

VARIATION RULES (必須のバリエーションルール):
🎲 NEVER repeat the same greeting twice
🎲 Randomly mix: やあ、こんにちは、よっ、ハーイ、おっす、へい
🎲 Vary particles: よ、ね、な、わ、さ、ぞ、だよ、です
🎲 Add emotions: ！、♪、～、☆、？
🎲 Change energy: energetic ↔ calm ↔ playful

絶対に同じ言葉を繰り返さない！毎回違う日本語で返答する！

[Response ID: {random.randint(10000, 99999)}]"""  # Random ID to break caching

    # Determine which mood to use based on mode
    if evil_mode:
        from utils.evil_mode import get_current_evil_mood, load_evil_mood_description
        current_mood_name, current_mood = get_current_evil_mood()
        is_sleeping = False
        forced_angry_until = None
        just_woken_up = False
        logger.info(f"Using Evil mode with mood: {current_mood_name}")
    else:
        current_mood = globals.DM_MOOD_DESCRIPTION  # Default to DM mood
        current_mood_name = globals.DM_MOOD  # Default to DM mood name
        is_sleeping = False  # DMs don't sleep
        forced_angry_until = None  # DMs don't get forced angry
        just_woken_up = False  # DMs don't wake up
    
    # If guild_id is provided, try to use server-specific mood (only in non-evil mode)
    if guild_id is not None and not evil_mode:
        try:
            from server_manager import server_manager
            server_config = server_manager.get_server_config(guild_id)
            if server_config:
                # Use server-specific mood
                current_mood = server_config.current_mood_description or load_mood_description(server_config.current_mood_name)
                current_mood_name = server_config.current_mood_name
                is_sleeping = server_config.is_sleeping
                forced_angry_until = server_config.forced_angry_until
                just_woken_up = server_config.just_woken_up
                logger.debug(f"Using server mood: {current_mood_name} for guild {guild_id}")
            else:
                logger.warning(f"No server config found for guild {guild_id}, using DM mood")
        except Exception as e:
            logger.error(f"Failed to get server mood for guild {guild_id}, falling back to DM mood: {e}")
            # Fall back to DM mood if server mood fails
    elif not evil_mode:
        logger.debug(f"Using DM mood: {globals.DM_MOOD}")

    # Append angry wake-up note if JUST_WOKEN_UP flag is set (only in non-evil mode)
    if just_woken_up and not evil_mode:
        system_prompt += (
            "\n\n🚨IMPORTANT NOTE🚨: Miku was just rudely woken up by the user with the current prompt of theirs.\nShe is fuming with anger and should respond harshly and coldly.\nDo NOT show kindness or friendliness in this response. Feel free to put certain words in all capital letters to emphasize your frustration."
        )


    # Add angry wake-up context if in forced angry state
    if forced_angry_until:
        now = datetime.datetime.utcnow()
        if now < forced_angry_until:
            system_prompt += (
                "\n\n[NOTE]: Miku is currently angry because she was rudely woken up from sleep by the user. "
                "Her responses should reflect irritation and coldness towards the user."
            )

    # Build conversation history - limit to prevent context overflow
    # 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: Modify system to understand Japanese mode
    # but DON'T add visible markers that waste tokens or get echoed
    # Instead, we rely on the strong system prompt to enforce Japanese
    
    # Add current user message (only if not empty)
    if user_prompt and user_prompt.strip():
        # Format with author name if provided (for server context)
        if author_name:
            content = f"{author_name}: {user_prompt}"
        else:
            content = user_prompt
        
        # Don't add visible markers - rely on system prompt enforcement instead
        # This prevents token waste and echo issues
        
        messages.append({"role": "user", "content": content})

    # Check if user is asking about profile picture and add context if needed
    pfp_context = ""
    try:
        from utils.pfp_context import is_asking_about_pfp, get_pfp_context_addition
        if user_prompt and is_asking_about_pfp(user_prompt):
            pfp_addition = get_pfp_context_addition()
            if pfp_addition:
                pfp_context = pfp_addition
    except Exception as e:
        # Silently fail if pfp context can't be retrieved
        pass

    # Combine structured prompt as a system message
    character_name = "Evil Miku" if evil_mode else "Miku"
    full_system_prompt = f"""{miku_context}

## CURRENT SITUATION
{character_name} is currently feeling: {current_mood}
Please respond in a way that reflects this emotional tone.{pfp_context}"""


    # Add media type awareness if provided
    if media_type:
        media_descriptions = {
            "image": "The user has sent you an image.",
            "video": "The user has sent you a video clip.",
            "gif": "The user has sent you an animated GIF.",
            "tenor_gif": "The user has sent you an animated GIF (from Tenor - likely a reaction GIF or meme)."
        }
        media_note = media_descriptions.get(media_type, f"The user has sent you {media_type}.")
        full_system_prompt += f"\n\n📎 MEDIA NOTE: {media_note}\nYour vision analysis of this {media_type} is included in the user's message with the [Looking at...] prefix."

    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":
        # Add random variation to temperature itself to prevent identical outputs
        base_temp = 1.1
        temp_variation = random.uniform(-0.1, 0.1)  # Random variation ±0.1
        temperature = base_temp + temp_variation
        
        top_p = 0.95
        frequency_penalty = 0.6  # Even stronger penalty
        presence_penalty = 0.6  # Even stronger encouragement for new content
        # Add random seed to ensure different responses each time
        seed = random.randint(0, 2**32 - 1)
        
        # Log the variation for debugging
        logger.debug(f"Japanese mode variation: temp={temperature:.2f}, seed={seed}")
    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": 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
            llama_url = get_current_gpu_url()
            logger.debug(f"Using GPU endpoint: {llama_url}")
            
            # Add timeout to prevent hanging indefinitely
            timeout = aiohttp.ClientTimeout(total=300)  # 300 second timeout
            async with session.post(f"{llama_url}/v1/chat/completions", json=payload, headers=headers, timeout=timeout) as response:
                if response.status == 200:
                    data = await response.json()
                    reply = data.get("choices", [{}])[0].get("message", {}).get("content", "No response.")
                    
                    # Strip surrounding quotes if present
                    reply = _strip_surrounding_quotes(reply)
                    
                    # Strip Japanese mode markers if in Japanese mode (prevent echo)
                    if globals.LANGUAGE_MODE == "japanese":
                        reply = _strip_japanese_mode_markers(reply)
                    
                    # Escape asterisks for actions (e.g., *adjusts hair* becomes \*adjusts hair\*)
                    reply = _escape_markdown_actions(reply)
                    
                    # Check if the reply is an error response and handle it
                    reply = await handle_response_error(
                        reply,
                        user_prompt=user_prompt,
                        user_id=str(user_id),
                        guild_id=str(guild_id) if guild_id else None,
                        author_name=author_name
                    )
                    
                    # Save to conversation history (only if both prompt and reply are non-empty)
                    # Don't save error messages to history
                    if user_prompt and user_prompt.strip() and reply and reply.strip() and reply != "Someone tell Koko-nii there is a problem with my AI.":
                        # Add user message to history
                        conversation_history.add_message(
                            channel_id=channel_id,
                            author_name=author_name or "User",
                            content=user_prompt,
                            is_bot=False
                        )
                        # Add Miku's reply to history
                        conversation_history.add_message(
                            channel_id=channel_id,
                            author_name="Miku",
                            content=reply,
                            is_bot=True
                        )
                    
                    # Also save to legacy globals for backward compatibility (skip error messages)
                    if user_prompt and user_prompt.strip() and reply and reply.strip() and reply != "Someone tell Koko-nii there is a problem with my AI.":
                        globals.conversation_history[user_id].append((user_prompt, reply))
                    
                    return reply
                else:
                    error_text = await response.text()
                    logger.error(f"Error from llama-swap: {response.status} - {error_text}")
                    
                    # Send webhook notification for HTTP errors
                    await handle_response_error(
                        f"Error: {response.status}",
                        user_prompt=user_prompt,
                        user_id=str(user_id),
                        guild_id=str(guild_id) if guild_id else None,
                        author_name=author_name
                    )
                    
                    # Don't save error responses to conversation history
                    return "Someone tell Koko-nii there is a problem with my AI."
        except asyncio.TimeoutError:
            logger.error("Timeout error in query_llama")
            return await handle_llm_error(
                asyncio.TimeoutError("Request timed out after 300 seconds"),
                user_prompt=user_prompt,
                user_id=str(user_id),
                guild_id=str(guild_id) if guild_id else None,
                author_name=author_name
            )
        except Exception as e:
            logger.error(f"Error in query_llama: {e}")
            return await handle_llm_error(
                e,
                user_prompt=user_prompt,
                user_id=str(user_id),
                guild_id=str(guild_id) if guild_id else None,
                author_name=author_name
            )

# Backward compatibility alias for existing code
query_ollama = query_llama