LLM.py

# LLM.py (V19.5 - Remove Bias Scores from Prompt)
import os, traceback, json, time, re
import httpx
from datetime import datetime
from typing import List, Dict, Any, Optional

# (استخدام مكتبة OpenAI الرسمية بدلاً من httpx)
from openai import AsyncOpenAI, RateLimitError, APIError

try:
    from r2 import R2Service
    from learning_hub.hub_manager import LearningHubManager # (استيراد العقل)
except ImportError:
    print("❌ [LLMService] فشل استيراد R2Service أو LearningHubManager")
    R2Service = None
    LearningHubManager = None

# (V8.1) استيراد NewsFetcher
try:
    from sentiment_news import NewsFetcher
except ImportError:
    NewsFetcher = None

# (استيراد VADER هنا أيضاً للـ type hinting)
try:
    from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer
except ImportError:
    SentimentIntensityAnalyzer = None


# (تحديث الإعدادات الافتراضية لتطابق NVIDIA)
LLM_API_URL = os.getenv("LLM_API_URL", "https://integrate.api.nvidia.com/v1")
LLM_API_KEY = os.getenv("LLM_API_KEY") # (هذا هو المفتاح الذي سيتم استخدامه)
LLM_MODEL = os.getenv("LLM_MODEL", "nvidia/llama-3.1-nemotron-ultra-253b-v1")

# (البارامترات المحددة من طرفك)
LLM_TEMPERATURE = 0.2
LLM_TOP_P = 0.7
LLM_MAX_TOKENS = 16384
LLM_FREQUENCY_PENALTY = 0.8
LLM_PRESENCE_PENALTY = 0.5

# إعدادات العميل
CLIENT_TIMEOUT = 300.0 

class LLMService:
    def __init__(self):
        if not LLM_API_KEY:
            raise ValueError("❌ [LLMService] متغير بيئة LLM_API_KEY غير موجود!")
        
        try:
            self.client = AsyncOpenAI(
                base_url=LLM_API_URL,
                api_key=LLM_API_KEY,
                timeout=CLIENT_TIMEOUT
            )
            # 🔴 --- START OF CHANGE (V19.5) --- 🔴
            print(f"✅ [LLMService V19.5] مهيأ. النموذج: {LLM_MODEL}")
            # 🔴 --- END OF CHANGE --- 🔴
            print(f"   -> Endpoint: {LLM_API_URL}")
        except Exception as e:
            # 🔴 --- START OF CHANGE (V19.5) --- 🔴
            print(f"❌ [LLMService V19.5] فشل تهيئة AsyncOpenAI: {e}")
            # 🔴 --- END OF CHANGE --- 🔴
            traceback.print_exc()
            raise
        
        # --- (الربط بالخدمات الأخرى) ---
        self.r2_service: Optional[R2Service] = None
        self.learning_hub: Optional[LearningHubManager] = None
        self.news_fetcher: Optional[NewsFetcher] = None
        self.vader_analyzer: Optional[SentimentIntensityAnalyzer] = None

    async def _call_llm(self, prompt: str) -> Optional[str]:
        """
        (محدث V19.2)
        إجراء استدعاء API للنموذج الضخم (يستخدم الآن "detailed thinking on" كـ system prompt).
        """
        
        system_prompt = "detailed thinking on"

        payload = {
            "model": LLM_MODEL,
            "messages": [
                {"role": "system", "content": system_prompt},
                {"role": "user", "content": prompt} # (prompt يحتوي الآن على تعليمات JSON)
            ],
            "temperature": LLM_TEMPERATURE,
            "top_p": LLM_TOP_P,
            "max_tokens": LLM_MAX_TOKENS,
            "frequency_penalty": LLM_FREQUENCY_PENALTY,
            "presence_penalty": LLM_PRESENCE_PENALTY,
            "stream": False, # (يجب أن تكون False للحصول على JSON)
            "response_format": {"type": "json_object"}
        }
        
        try:
            response = await self.client.chat.completions.create(**payload)
            
            if response.choices and len(response.choices) > 0:
                content = response.choices[0].message.content
                if content:
                    return content.strip()
            
            print(f"❌ [LLMService] استجابة API غير متوقعة: {response.model_dump_json()}")
            return None
            
        except RateLimitError as e:
            print(f"❌ [LLMService] خطأ Rate Limit من NVIDIA API: {e}")
        except APIError as e:
            print(f"❌ [LLMService] خطأ API من NVIDIA API: {e}")
        except json.JSONDecodeError:
            print(f"❌ [LLMService] فشل في تحليل استجابة JSON.")
        except Exception as e:
            print(f"❌ [LLMService] خطأ غير متوقع في _call_llm: {e}")
            traceback.print_exc()
            
        return None

    def _parse_llm_response_enhanced(self, 
                                     response_text: str, 
                                     fallback_strategy: str = "decision", 
                                     symbol: str = "N/A") -> Optional[Dict[str, Any]]:
        """
        (محدث V8) محلل JSON ذكي ومتسامح مع الأخطاء.
        """
        if not response_text:
            print(f"   ⚠️ [LLMParser] الاستجابة فارغة لـ {symbol}.")
            return self._get_fallback_response(fallback_strategy, "Empty response")

        # 1. محاولة تحليل JSON مباشرة (لأننا طلبنا response_format=json_object)
        try:
            return json.loads(response_text)
        except json.JSONDecodeError:
            print(f"   ⚠️ [LLMParser] فشل تحليل JSON المباشر لـ {symbol}. محاولة استخراج JSON...")
            pass # (الانتقال إلى المحاولة 2)

        # 2. محاولة استخراج JSON من داخل نص (Fallback 1)
        try:
            # (البحث عن أول { وآخر })
            json_match = re.search(r'\{.*\}', response_text, re.DOTALL)
            if json_match:
                json_string = json_match.group(0)
                return json.loads(json_string)
            else:
                print(f"   ⚠️ [LLMParser] لم يتم العثور على JSON مطابق لـ {symbol}.")
                raise json.JSONDecodeError("No JSON object found in text", response_text, 0)
        except json.JSONDecodeError as e:
            print(f"   ❌ [LLMParser] فشل الاستخراج النهائي لـ {symbol}. نص الاستجابة: {response_text[:200]}...")
            return self._get_fallback_response(fallback_strategy, f"Final JSON parse fail: {e}")
        except Exception as e:
            print(f"   ❌ [LLMParser] خطأ غير متوقع في المحلل لـ {symbol}: {e}")
            return self._get_fallback_response(fallback_strategy, f"Unexpected parser error: {e}")

    def _get_fallback_response(self, strategy: str, reason: str) -> Optional[Dict[str, Any]]:
        """
        (محدث V8) إرجاع استجابة آمنة عند فشل النموذج الضخم.
        """
        print(f"   🚨 [LLMService] تفعيل الاستجابة الاحتياطية (Fallback) لاستراتيجية '{strategy}' (السبب: {reason})")
        
        if strategy == "decision":
            # (القرار الآمن: لا تتداول)
            return {
                "action": "NO_DECISION",
                "strategy_to_watch": "GENERIC",
                "confidence_level": 0,
                "reasoning": f"LLM analysis failed: {reason}",
                "exit_profile": "Standard"
            }
        elif strategy == "reanalysis":
             # (القرار الآمن: استمر في الصفقة الحالية)
            return {
                "action": "HOLD",
                "strategy": "MAINTAIN_CURRENT",
                "reasoning": f"LLM re-analysis failed: {reason}. Maintaining current trade strategy."
            }
        elif strategy == "reflection":
            # (القرار الآمن: لا تقم بإنشاء قاعدة تعلم)
            return None # (سيمنع Reflector من إنشاء دلتا)
            
        elif strategy == "distillation":
             # (القرار الآمن: لا تقم بإنشاء قواعد مقطرة)
            return None # (سيمنع Curator من المتابعة)
            
        return None # (Fallback عام)

    async def get_trading_decision(self, candidate_data: Dict[str, Any]) -> Optional[Dict[str, Any]]:
        """
        (محدث V8.1)
        يستدعي النموذج الضخم لاتخاذ قرار "WATCH" استراتيجي (Explorer Brain).
        """
        symbol = candidate_data.get('symbol', 'UNKNOWN')
        try:
            # 1. (العقل) جلب القواعد (Deltas) من محور التعلم
            learning_context_prompt = "Playbook: No learning context available."
            if self.learning_hub:
                learning_context_prompt = await self.learning_hub.get_active_context_for_llm(
                    domain="general", 
                    query=f"{symbol} strategy decision"
                )
            
            # 2. إنشاء الـ Prompt (باللغة الإنجليزية)
            prompt = self._create_trading_prompt(candidate_data, learning_context_prompt)
            
            if self.r2_service:
                await self.r2_service.save_llm_prompts_async(symbol, "trading_decision", prompt, candidate_data)

            # 3. استدعاء النموذج الضخم (LLM)
            response_text = await self._call_llm(prompt)

            # 4. تحليل الاستجابة (باستخدام المحلل الذكي)
            decision_json = self._parse_llm_response_enhanced(
                response_text, 
                fallback_strategy="decision", 
                symbol=symbol
            )
            
            return decision_json

        except Exception as e:
            print(f"❌ [LLMService] فشل فادح في get_trading_decision لـ {symbol}: {e}")
            traceback.print_exc()
            return self._get_fallback_response("decision", str(e)) # (إرجاع قرار آمن)

    async def re_analyze_trade_async(self, trade_data: Dict[str, Any], current_data: Dict[str, Any]) -> Optional[Dict[str, Any]]:
        """
        (محدث V19.3)
        يستدعي النموذج الضخم لإعادة تحليل صفقة مفتوحة (Reflector Brain).
        """
        symbol = trade_data.get('symbol', 'UNKNOWN')
        try:
            # 1. (العقل) جلب القواعد (Deltas) من محور التعلم
            learning_context_prompt = "Playbook: No learning context available."
            if self.learning_hub:
                learning_context_prompt = await self.learning_hub.get_active_context_for_llm(
                    domain="strategy", 
                    query=f"{symbol} re-analysis {trade_data.get('strategy', 'GENERIC')}"
                )

            # 2. (V8.1) جلب أحدث الأخبار (باستخدام NewsFetcher المخصص)
            latest_news_text = "News data unavailable for re-analysis."
            latest_news_score = 0.0
            
            # (استخدام self.vader_analyzer الذي تم حقنه)
            if self.news_fetcher:
                latest_news_text = await self.news_fetcher.get_news_for_symbol(symbol)
                if self.vader_analyzer and latest_news_text: # (التحقق من المحلل المُمرر)
                     vader_scores = self.vader_analyzer.polarity_scores(latest_news_text)
                     latest_news_score = vader_scores.get('compound', 0.0)
            
            current_data['latest_news_text'] = latest_news_text
            current_data['latest_news_score'] = latest_news_score

            # 3. إنشاء الـ Prompt (باللغة الإنجليزية)
            prompt = await self._create_reanalysis_prompt(trade_data, current_data, learning_context_prompt)
            
            if self.r2_service:
                 await self.r2_service.save_llm_prompts_async(symbol, "trade_reanalysis", prompt, current_data)

            # 4. استدعاء النموذج الضخم (LLM)
            response_text = await self._call_llm(prompt)

            # 5. تحليل الاستجابة (باستخدام المحلل الذكي)
            decision_json = self._parse_llm_response_enhanced(
                response_text, 
                fallback_strategy="reanalysis", 
                symbol=symbol
            )
            
            return decision_json

        except Exception as e:
            print(f"❌ [LLMService] فشل فادح في re_analyze_trade_async لـ {symbol}: {e}")
            traceback.print_exc()
            return self._get_fallback_response("reanalysis", str(e)) # (إرجاع قرار آمن)
            
    # --- (دوال إنشاء الـ Prompts) ---
    # (ملاحظة: هذه الدوال يجب أن تكون دائماً باللغة الإنجليزية)

    def _create_trading_prompt(self, 
                               candidate_data: Dict[str, Any], 
                               learning_context: str) -> str:
        """
        (معدل V19.5)
        إنشاء الـ Prompt (باللغة الإنجليزية) لاتخاذ قرار التداول الأولي (Explorer).
        (تمت إزالة الدرجات المسبقة لتقليل الانحياز)
        """
        
        symbol = candidate_data.get('symbol', 'N/A')
        
        # --- 1. استخراج بيانات ML (الطبقة 1) ---
        # (تمت إزالة l1_score و l1_reasons عمداً)
        pattern_data = candidate_data.get('pattern_analysis', {})
        mc_data = candidate_data.get('monte_carlo_distribution', {})
        
        # --- 2. استخراج بيانات المشاعر والأخبار (الطبقة 1) ---
        news_text = candidate_data.get('news_text', 'No news text provided.')
        news_score_raw = candidate_data.get('news_score_raw', 0.0)
        statistical_news_pnl = candidate_data.get('statistical_news_pnl', 0.0)
        
        # --- 3. استخراج بيانات الحيتان (الطبقة 1) ---
        whale_data = candidate_data.get('whale_data', {})
        whale_summary = whale_data.get('llm_friendly_summary', {})
        exchange_flows = whale_data.get('exchange_flows', {})
        
        whale_signal = whale_summary.get('recommended_action', 'HOLD')
        whale_confidence = whale_summary.get('confidence', 0.3)
        whale_reason = whale_summary.get('whale_activity_summary', 'No whale data.')
        net_flow_usd = exchange_flows.get('net_flow_usd', 0.0)

        # (البيانات طويلة المدى - من تحليل 24 ساعة الجديد)
        accumulation_data_24h = whale_data.get('accumulation_analysis_24h', {})
        net_flow_24h_usd = accumulation_data_24h.get('net_flow_usd', 0.0)
        total_inflow_24h_usd = accumulation_data_24h.get('to_exchanges_usd', 0.0)
        total_outflow_24h_usd = accumulation_data_24h.get('from_exchanges_usd', 0.0)
        relative_net_flow_24h_percent = accumulation_data_24h.get('relative_net_flow_percent', 0.0)

        # --- 4. استخراج بيانات السوق (الطبقة 0) ---
        market_context = candidate_data.get('sentiment_data', {})
        market_trend = market_context.get('market_trend', 'UNKNOWN')
        btc_sentiment = market_context.get('btc_sentiment', 'UNKNOWN')
        
        # --- 5. بناء أقسام الـ Prompt (الإنجليزية) ---
        
        playbook_prompt = f"""
--- START OF LEARNING PLAYBOOK ---
{learning_context}
--- END OF PLAYBOOK ---
"""
        
        # 🔴 --- START OF CHANGE (V19.5) --- 🔴
        # (تمت إزالة درجة l1_score و l1_reasons من هنا)
        technical_summary_prompt = f"""
1.  **Technical Analysis:**
    * Chart Pattern: {pattern_data.get('pattern_detected', 'None')} (Conf: {pattern_data.get('pattern_confidence', 0):.2f})
    * Monte Carlo (1h): {mc_data.get('probability_of_gain', 0) * 100:.1f}% chance of profit (Expected: {mc_data.get('expected_return_pct', 0):.2f}%)
"""
        # 🔴 --- END OF CHANGE --- 🔴

        news_prompt = f"""
2.  **News & Sentiment Analysis:**
    * Market Trend: {market_trend} (BTC: {btc_sentiment})
    * VADER (Raw): {news_score_raw:.3f}
    * Statistical PnL (Learned): {statistical_news_pnl:+.2f}%
    * News Text: {news_text[:300]}...
"""
        whale_activity_prompt = f"""
3.  **Whale Activity (Real-time Flow - Optimized Window):**
    * Signal: {whale_signal} (Confidence: {whale_confidence:.2f})
    * Reason: {whale_reason}
    * Net Flow (to/from Exchanges): ${net_flow_usd:,.2f}

4.  **Whale Activity (24h Accumulation):**
    * 24h Net Flow (Accumulation): ${net_flow_24h_usd:,.2f}
    * 24h Total Deposits: ${total_inflow_24h_usd:,.2f}
    * 24h Total Withdrawals: ${total_outflow_24h_usd:,.2f}
    * Relative 24h Net Flow (vs Daily Volume): {relative_net_flow_24h_percent:+.2f}%
"""
        
        # 🔴 --- START OF CHANGE (V19.5) --- 🔴
        # (تم تحديث التعليمات ليعكس تحليل البيانات "الخام" بدلاً من الدرجات)
        task_prompt = f"""
CONTEXT:
You are an expert AI trading analyst (Explorer Brain). 
Analyze the following raw technical, news, and whale data for {symbol}. You must make a decision based *only* on the data provided, without any pre-calculated scores.
Decide if this combination of signals presents a high-potential opportunity to 'WATCH'.
{playbook_prompt}

--- START OF CANDIDATE DATA ---
{technical_summary_prompt}
{news_prompt}
{whale_activity_prompt}
--- END OF CANDIDATE DATA ---

TASK:
1.  **Internal Thinking (Private):** Perform a step-by-step analysis (as triggered by the system prompt).
    * Synthesize all data points (Chart Pattern, Monte Carlo, News, Whale Flow, 24h Accumulation).
    * Are the signals aligned? (e.g., Bullish Pattern + High MC Probability + Whale Accumulation = Strong).
    * Are there conflicts? (e.g., Bullish Pattern but high 24h Deposits = Risky).
    * Consult the "Playbook" for learned rules.
2.  **Final Decision:** Based on your internal thinking, decide the final action.
3.  **Output Constraint:** Provide your final answer ONLY in the requested JSON object format, with no introductory text, markdown formatting, or explanations.

OUTPUT (JSON Object ONLY):
{{
  "action": "WATCH" or "NO_DECISION",
  "strategy_to_watch": "STRATEGY_NAME",
  "confidence_level": 0.0_to_1.0,
  "reasoning": "Brief justification (max 40 words) synthesizing all data points.",
  "exit_profile": "Aggressive" or "Standard" or "Patient"
}}
"""
        # 🔴 --- END OF CHANGE --- 🔴
        
        # (نرسل فقط task_prompt لأنه يحتوي الآن على كل شيء)
        return task_prompt


    async def _create_reanalysis_prompt(self, 
                                    trade_data: Dict[str, Any], 
                                    current_data: Dict[str, Any], 
                                    learning_context: str) -> str:
        """
        (معدل V19.4)
        إنشاء الـ Prompt (باللغة الإنجليزية) لإعادة تحليل صفقة مفتوحة (Reflector Brain).
        (تم إصلاح تنسيق مونت كارلو)
        """
        
        symbol = trade_data.get('symbol', 'N/A')
        
        # --- 1. بيانات الصفقة الأصلية (القديمة) ---
        original_strategy = trade_data.get('strategy', 'N/A')
        original_reasoning = trade_data.get('decision_data', {}).get('reasoning', 'N/A')
        entry_price = trade_data.get('entry_price', 0)
        current_pnl = trade_data.get('pnl_percent', 0)
        current_sl = trade_data.get('stop_loss', 0)
        current_tp = trade_data.get('take_profit', 0)
        
        # --- 2. البيانات الفنية المحدثة (الحالية) ---
        current_price = current_data.get('current_price', 0)
        mc_data = current_data.get('monte_carlo_distribution', {})
        mc_prob = mc_data.get('probability_of_gain', 0)
        mc_expected_return = mc_data.get('expected_return_pct', 0)
        
        # --- 3. (V8.1) بيانات الأخبار المحدثة (الحالية) ---
        latest_news_text = current_data.get('latest_news_text', 'No news.')
        latest_news_score = current_data.get('latest_news_score', 0.0)

        # --- 4. (العقل) بيانات التعلم الإحصائي ---
        statistical_feedback = ""
        if self.learning_hub:
            statistical_feedback = await self.learning_hub.get_statistical_feedback_for_llm(original_strategy)

        # --- 5. بناء أقسام الـ Prompt (الإنجليزية) ---
        
        playbook_prompt = f"""
--- START OF LEARNING PLAYBOOK ---
{learning_context}
{statistical_feedback}
--- END OF PLAYBOOK ---
"""

        trade_status_prompt = f"""
1.  **Open Trade Status ({symbol}):**
    * Current PnL: {current_pnl:+.2f}%
    * Original Strategy: {original_strategy}
    * Original Reasoning: {original_reasoning}
    * Entry Price: {entry_price}
    * Current Price: {current_price}
    * Current StopLoss: {current_sl}
    * Current TakeProfit: {current_tp}
"""

        current_analysis_prompt = f"""
2.  **Current Real-time Analysis:**
    * Monte Carlo (1h): {mc_prob * 100:.1f}% chance of profit (Expected: {mc_expected_return:.2f}%)
    * Latest News (VADER: {latest_news_score:.3f}): {latest_news_text[:300]}...
"""
        
        # (دمج جميع التعليمات في رسالة الـ user)
        task_prompt = f"""
CONTEXT:
You are an expert AI trading analyst (Reflector Brain). 
An open trade for {symbol} has triggered a mandatory re-analysis. Analyze the new data and decide the next action.
{playbook_prompt}

--- START OF TRADE DATA ---
{trade_status_prompt}
{current_analysis_prompt}
--- END OF TRADE DATA ---

TASK:
1.  **Internal Thinking (Private):** Perform a step-by-step analysis (as triggered by the system prompt).
    * Compare the "Open Trade Status" with the "Current Real-time Analysis".
    * Has the situation improved or deteriorated? (e.g., PnL is good, but new Monte Carlo is negative).
    * Are there new critical news?
    * Consult the "Playbook" for learned rules and statistical feedback.
2.  **Final Decision:** Based on your internal thinking, decide the best course of action (HOLD, UPDATE_TRADE, CLOSE_TRADE).
3.  **Output Constraint:** Provide your final answer ONLY in the requested JSON object format, with no introductory text, markdown formatting, or explanations.

OUTPUT (JSON Object ONLY):
{{
  "action": "HOLD" or "UPDATE_TRADE" or "CLOSE_TRADE",
  "strategy": "MAINTAIN_CURRENT" or "ADAPTIVE_EXIT" or "IMMEDIATE_EXIT",
  "reasoning": "Brief justification (max 40 words) for the decision.",
  "new_stop_loss": (float or null, required if action is 'UPDATE_TRADE'),
  "new_take_profit": (float or null, required if action is 'UPDATE_TRADE')
}}
"""
        
        return task_prompt