Update LLM.py

LLM.py

@@ -1,9 +1,12 @@
-# LLM.py (V18.2 - Async Fix)
+# LLM.py (V19.2 - NVIDIA Engine + System Trigger Fix)
 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)
@@ -18,13 +21,19 @@ try:
 except ImportError:
     NewsFetcher = None
 
-# إعدادات النموذج
-LLM_API_URL = os.getenv("LLM_API_URL", "https://api.groq.com/openai/v1/chat/completions")
-LLM_API_KEY = os.getenv("LLM_API_KEY")
-LLM_MODEL = os.getenv("LLM_MODEL", "llama3-70b-8192")
+# (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
-# (زيادة المهلة إلى 5 دقائق للتحليلات المعقدة)
 CLIENT_TIMEOUT = 300.0 
 
 class LLMService:
@@ -32,58 +41,68 @@ class LLMService:
         if not LLM_API_KEY:
             raise ValueError("❌ [LLMService] متغير بيئة LLM_API_KEY غير موجود!")
         
-        self.headers = {
-            "Authorization": f"Bearer {LLM_API_KEY}",
-            "Content-Type": "application/json"
-        }
-        self.http_client = httpx.AsyncClient(timeout=CLIENT_TIMEOUT)
+        try:
+            self.client = AsyncOpenAI(
+                base_url=LLM_API_URL,
+                api_key=LLM_API_KEY,
+                timeout=CLIENT_TIMEOUT
+            )
+            print(f"✅ [LLMService V19.2] مهيأ. النموذج: {LLM_MODEL}")
+            print(f"   -> Endpoint: {LLM_API_URL}")
+        except Exception as e:
+            print(f"❌ [LLMService V19.2] فشل تهيئة AsyncOpenAI: {e}")
+            traceback.print_exc()
+            raise
         
         # --- (الربط بالخدمات الأخرى) ---
         self.r2_service: Optional[R2Service] = None
         self.learning_hub: Optional[LearningHubManager] = None
-        
-        # (V8.1) إضافة NewsFetcher (للاستخدام في إعادة التحليل)
         self.news_fetcher: Optional[NewsFetcher] = None
-        
-        print(f"✅ [LLMService] مهيأ. النموذج: {LLM_MODEL}")
 
-    async def _call_llm(self, prompt: str, temperature: float = 0.2, max_tokens: int = 1500) -> Optional[str]:
+    async def _call_llm(self, prompt: str) -> Optional[str]:
         """
-        (محدث) إجراء استدعاء API للنموذج الضخم (باللغة الإنجليزية).
+        (محدث V19.2)
+        إجراء استدعاء API للنموذج الضخم (يستخدم الآن "detailed thinking on" كـ system prompt).
         """
+        
+        # 🔴 --- START OF CHANGE (V19.2) --- 🔴
+        # (استخدام "detailed thinking on" كـ system prompt كما طلبت)
+        system_prompt = "detailed thinking on"
+        # (تم نقل جميع التعليمات الأخرى إلى الـ User Prompt)
+        # 🔴 --- END OF CHANGE --- 🔴
+
         payload = {
             "model": LLM_MODEL,
             "messages": [
-                {"role": "system", "content": "You are an expert AI trading analyst. You MUST respond ONLY with the requested JSON object, with no introductory text, markdown formatting (like ```json), or explanations."},
-                {"role": "user", "content": prompt}
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": prompt} # (prompt يحتوي الآن على تعليمات JSON)
             ],
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-            "top_p": 1,
-            "stream": False,
+            "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.http_client.post(LLM_API_URL, headers=self.headers, json=payload)
-            response.raise_for_status() # Check for HTTP 4xx/5xx errors
-            
-            data = response.json()
+            response = await self.client.chat.completions.create(**payload)
             
-            if "choices" in data and len(data["choices"]) > 0:
-                content = data["choices"][0].get("message", {}).get("content")
+            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: {data}")
+            print(f"❌ [LLMService] استجابة API غير متوقعة: {response.model_dump_json()}")
             return None
             
-        except httpx.HTTPStatusError as http_err:
-            print(f"❌ [LLMService] خطأ HTTP: {http_err} - {http_err.response.text}")
-        except httpx.RequestError as req_err:
-            print(f"❌ [LLMService] خطأ في الطلب: {req_err}")
+        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: {response.text}")
+            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()
@@ -165,10 +184,8 @@ class LLMService:
         symbol = candidate_data.get('symbol', 'UNKNOWN')
         try:
             # 1. (العقل) جلب القواعد (Deltas) من محور التعلم
-            # (سيتم جلب القواعد العامة + قواعد الاستراتيجية بناءً على المرشح)
             learning_context_prompt = "Playbook: No learning context available."
             if self.learning_hub:
-                # (استعلام عام لجلب أفضل القواعد الشاملة)
                 learning_context_prompt = await self.learning_hub.get_active_context_for_llm(
                     domain="general", 
                     query=f"{symbol} strategy decision"
@@ -177,12 +194,11 @@ class LLMService:
             # 2. إنشاء الـ Prompt (باللغة الإنجليزية)
             prompt = self._create_trading_prompt(candidate_data, learning_context_prompt)
             
-            # (اختياري: حفظ الـ Prompt للتدقيق)
             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, temperature=0.1, max_tokens=1000)
+            response_text = await self._call_llm(prompt)
 
             # 4. تحليل الاستجابة (باستخدام المحلل الذكي)
             decision_json = self._parse_llm_response_enhanced(
@@ -208,7 +224,6 @@ class LLMService:
             # 1. (العقل) جلب القواعد (Deltas) من محور التعلم
             learning_context_prompt = "Playbook: No learning context available."
             if self.learning_hub:
-                # (استعلام محدد لإعادة التحليل)
                 learning_context_prompt = await self.learning_hub.get_active_context_for_llm(
                     domain="strategy", 
                     query=f"{symbol} re-analysis {trade_data.get('strategy', 'GENERIC')}"
@@ -218,28 +233,22 @@ class LLMService:
             latest_news_text = "News data unavailable for re-analysis."
             latest_news_score = 0.0
             if self.news_fetcher:
-                # (هذا يجلب أحدث الأخبار في الوقت الفعلي)
                 latest_news_text = await self.news_fetcher.get_news_for_symbol(symbol)
                 if self.news_fetcher.vader_analyzer and latest_news_text:
                      vader_scores = self.news_fetcher.vader_analyzer.polarity_scores(latest_news_text)
                      latest_news_score = vader_scores.get('compound', 0.0)
             
-            # (إضافة الأخبار المحدثة إلى البيانات الحالية)
             current_data['latest_news_text'] = latest_news_text
             current_data['latest_news_score'] = latest_news_score
 
             # 3. إنشاء الـ Prompt (باللغة الإنجليزية)
-            # 🔴 --- START OF CHANGE (V18.2 - Async Fix) --- 🔴
-            # (يجب استخدام await لأن الدالة أصبحت async)
             prompt = await self._create_reanalysis_prompt(trade_data, current_data, learning_context_prompt)
-            # 🔴 --- END OF CHANGE --- 🔴
             
-            # (اختياري: حفظ الـ Prompt للتدقيق)
             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, temperature=0.0, max_tokens=1000) # (درجة حرارة 0.0 للحسم)
+            response_text = await self._call_llm(prompt)
 
             # 5. تحليل الاستجابة (باستخدام المحلل الذكي)
             decision_json = self._parse_llm_response_enhanced(
@@ -262,8 +271,9 @@ class LLMService:
                                candidate_data: Dict[str, Any], 
                                learning_context: str) -> str:
         """
-        (معدل V18.1)
+        (معدل V19.2)
         إنشاء الـ Prompt (باللغة الإنجليزية) لاتخاذ قرار التداول الأولي (Explorer).
+        (تم نقل جميع التعليمات إلى هنا لتناسب system prompt الجديد)
         """
         
         symbol = candidate_data.get('symbol', 'N/A')
@@ -275,13 +285,11 @@ class LLMService:
         mc_data = candidate_data.get('monte_carlo_distribution', {})
         
         # --- 2. استخراج بيانات المشاعر والأخبار (الطبقة 1) ---
-        # (V8.2) جلب بيانات VADER الإحصائية (المتعلمة)
         news_text = candidate_data.get('news_text', 'No news text provided.')
-        news_score_raw = candidate_data.get('news_score_raw', 0.0) # (درجة VADER الخام)
-        statistical_news_pnl = candidate_data.get('statistical_news_pnl', 0.0) # (الدرجة الإحصائية المتعلمة)
+        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) ---
-        # (ملاحظة: هذا هو القسم الذي سنقوم بتحديثه)
         whale_data = candidate_data.get('whale_data', {})
         whale_summary = whale_data.get('llm_friendly_summary', {})
         exchange_flows = whale_data.get('exchange_flows', {})
@@ -289,18 +297,14 @@ class LLMService:
         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.')
-        
-        # (البيانات قصيرة المدى - نفترض أنها من أفضل نافذة متعلمة، e.g., 1h)
         net_flow_usd = exchange_flows.get('net_flow_usd', 0.0)
 
-        # 🔴 --- START OF CHANGE (V18.1) --- 🔴
         # (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)
-        # 🔴 --- END OF CHANGE --- 🔴
 
         # --- 4. استخراج بيانات السوق (الطبقة 0) ---
         market_context = candidate_data.get('sentiment_data', {})
@@ -309,20 +313,17 @@ class LLMService:
         
         # --- 5. بناء أقسام الـ Prompt (الإنجليزية) ---
         
-        # (التعلم)
         playbook_prompt = f"""
 --- START OF LEARNING PLAYBOOK ---
 {learning_context}
 --- END OF PLAYBOOK ---
 """
-        # (ملخص ML)
         ml_summary_prompt = f"""
 1.  **ML Analysis (Score: {l1_score:.3f}):**
     * Reasons: {', '.join(l1_reasons)}
     * 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):.1f}% chance of profit (Expected: {mc_data.get('expected_return_pct', 0):.2f}%)
 """
-        # (ملخص الأخبار)
         news_prompt = f"""
 2.  **News & Sentiment Analysis:**
     * Market Trend: {market_trend} (BTC: {btc_sentiment})
@@ -330,7 +331,6 @@ class LLMService:
     * Statistical PnL (Learned): {statistical_news_pnl:+.2f}%
     * News Text: {news_text[:300]}...
 """
-        # (ملخص الحيتان - محدث)
         whale_activity_prompt = f"""
 3.  **Whale Activity (Real-time Flow - Optimized Window):**
     * Signal: {whale_signal} (Confidence: {whale_confidence:.2f})
@@ -344,43 +344,50 @@ class LLMService:
     * Relative 24h Net Flow (vs Daily Volume): {relative_net_flow_24h_percent:+.2f}%
 """
         
-        prompt_sections = [
-            f"SYSTEM: You are an expert AI trading analyst (Explorer Brain). Analyze the provided data for {symbol} and decide if it's a high-potential candidate to 'WATCH'.",
-            playbook_prompt,
-            "--- START OF CANDIDATE DATA ---",
-            ml_summary_prompt,
-            news_prompt,
-            whale_activity_prompt,
-            "--- END OF CANDIDATE DATA ---",
-            """
+        # 🔴 --- START OF CHANGE (V19.2) --- 🔴
+        # (تم دمج جميع التعليمات هنا في رسالة الـ user)
+        task_prompt = f"""
+CONTEXT:
+You are an expert AI trading analyst (Explorer Brain). 
+Analyze the provided data for {symbol} and decide if it's a high-potential candidate to 'WATCH'.
+{playbook_prompt}
+
+--- START OF CANDIDATE DATA ---
+{ml_summary_prompt}
+{news_prompt}
+{whale_activity_prompt}
+--- END OF CANDIDATE DATA ---
+
 TASK:
-1.  Synthesize all data points (ML, News, Whale Flow, 24h Accumulation).
-2.  Consult the "Playbook" for learned rules.
-3.  Decide the final action: 'WATCH' (if high confidence) or 'NO_DECISION' (if low confidence or conflicting signals).
-4.  If 'WATCH', define the 'strategy_to_watch' (e.g., 'Breakout', 'MeanReversion', 'WhaleAccumulation') and an appropriate 'exit_profile' (e.g., 'Aggressive', 'Standard', 'Patient').
+1.  **Internal Thinking (Private):** Perform a step-by-step analysis (as triggered by the system prompt).
+    * Synthesize all data points (ML, News, Whale Flow, 24h Accumulation).
+    * Are the signals aligned? (e.g., ML breakout + Whale Accumulation = Strong).
+    * Are there conflicts? (e.g., Good ML Score 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 --- 🔴
         
-        return "\n".join(prompt_sections)
+        # (نرسل فقط task_prompt لأنه يحتوي الآن على كل شيء)
+        return task_prompt
+
 
-    # 🔴 --- START OF CHANGE (V18.2 - Async Fix) --- 🔴
-    # (يجب أن تكون الدالة async)
     async def _create_reanalysis_prompt(self, 
                                     trade_data: Dict[str, Any], 
                                     current_data: Dict[str, Any], 
                                     learning_context: str) -> str:
-    # 🔴 --- END OF CHANGE --- 🔴
         """
-        (معدل V8.1)
+        (معدل V19.2)
         إنشاء الـ Prompt (باللغة الإنجليزية) لإعادة تحليل صفقة مفتوحة (Reflector Brain).
         """
         
@@ -407,10 +414,7 @@ OUTPUT (JSON Object ONLY):
         # --- 4. (العقل) بيانات التعلم الإحصائي ---
         statistical_feedback = ""
         if self.learning_hub:
-            # 🔴 --- START OF CHANGE (V18.2 - Async Fix) --- 🔴
-            # (هذا هو السطر الذي تسبب بالخطأ - يجب استخدام await)
             statistical_feedback = await self.learning_hub.get_statistical_feedback_for_llm(original_strategy)
-            # 🔴 --- END OF CHANGE --- 🔴
 
         # --- 5. بناء أقسام الـ Prompt (الإنجليزية) ---
         
@@ -438,32 +442,37 @@ OUTPUT (JSON Object ONLY):
     * Latest News (VADER: {latest_news_score:.3f}): {latest_news_text[:300]}...
 """
         
-        prompt_sections = [
-            f"SYSTEM: 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 ---",
-            """
+        # 🔴 --- START OF CHANGE (V19.2) --- 🔴
+        # (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.  Analyze the "Open Trade Status" against the "Current Real-time Analysis".
-2.  Consult the "Playbook" for learned rules and statistical feedback.
-3.  Decide the best course of action:
-    * 'HOLD': Maintain the current strategy (if analysis is still valid or neutral).
-    * 'UPDATE_TRADE': Adjust StopLoss/TakeProfit (if risk/reward changed).
-    * 'CLOSE_TRADE': Immediately close the trade (if analysis is invalidated or high risk).
-4.  You MUST provide new SL/TP targets if action is 'UPDATE_TRADE'.
+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')
-}
+}}
 """
-        ]
+        # 🔴 --- END OF CHANGE --- 🔴
         
-        return "\n".join(prompt_sections)
+        return task_prompt