Update app.py

app.py

@@ -1,3 +1,4 @@
+
 import os
 import requests
 from datetime import datetime, timedelta, timezone
@@ -5,58 +6,38 @@ from flask import Flask, request, Response, stream_with_context, render_template
 
 app = Flask(__name__)
 
-# 🔐 --- SECURE ENVIRONMENT VARIABLES (Hugging Face Secrets) ---
+# 🔐 --- SECURE ENVIRONMENT VARIABLES ---
 API_KEY = os.environ.get("NVIDIA_API_KEY") or os.environ.get("YOUR_VEDIKA_API_KEY")
-MODEL_ID = os.environ.get("MODEL_ID") # Default fallback
-SERPAPI_KEY = os.environ.get("SERPAPI_KEY") # 🔒 Now secured in environment variables
+MODEL_ID = os.environ.get("MODEL_ID", "google/diffusiongemma-26b-a4b-it") 
+SERPAPI_KEY = os.environ.get("SERPAPI_KEY") # 🔒 Secured in Hugging Face Secrets
 
 # NVIDIA's official invoke URL
 INVOKE_URL = "https://integrate.api.nvidia.com/v1/chat/completions"
 
-# 🌐 --- SERPAPI GOOGLE SEARCH ENGINE (100% BULLETPROOF) --- 🌐
+# 🌐 --- SERPAPI GOOGLE SEARCH ENGINE ---
 def web_search_scraper(query, num_results=5):
-    """
-    यह SerpApi का उपयोग करके सीधे Google Search से एकदम सटीक और ताज़ा (JSON) डेटा लाता है।
-    """
     results = []
     if not SERPAPI_KEY:
-        print("SYSTEM ALERT: SERPAPI_KEY is missing in environment secrets.")
+        print("SYSTEM ALERT: SERPAPI_KEY missing in secrets.")
         return results
 
     try:
-        params = {
-            "engine": "google",
-            "q": query,
-            "api_key": SERPAPI_KEY,
-            "num": num_results,
-            "hl": "en",
-            "gl": "in" # India Region for local context
-        }
-        
+        params = {"engine": "google", "q": query, "api_key": SERPAPI_KEY, "num": num_results, "hl": "en", "gl": "in"}
         response = requests.get("https://serpapi.com/search", params=params, timeout=10)
         data = response.json()
         
-        # 'organic_results' से असली Google सर्च का डेटा निकालना
         if "organic_results" in data:
             for item in data["organic_results"]:
                 title = item.get("title", "")
                 link = item.get("link", "")
                 snippet = item.get("snippet", "")
-                
                 if title and snippet:
-                    results.append({
-                        "title": title, 
-                        "link": link, 
-                        "snippet": snippet
-                    })
-                    
+                    results.append({"title": title, "link": link, "snippet": snippet})
     except Exception as e:
         print(f"SerpApi Error: {e}")
 
     return results
 
-# ----------------------------------------------------
-
 @app.route('/')
 def home():
     try:
@@ -68,7 +49,7 @@ def home():
 @app.route('/api/chat', methods=['POST'])
 def chat():
     if not API_KEY or not INVOKE_URL or not MODEL_ID:
-        return Response("Server Error: Secrets missing. Check NVIDIA API Key.", status=500)
+        return Response("Server Error: Secrets missing.", status=500)
 
     data = request.get_json() or {}
     user_message = data.get("message", "")
@@ -78,11 +59,11 @@ def chat():
     max_tokens = data.get("max_tokens", 4096)
     temperature = data.get("temperature", 1.0)
     
-    # 🧠 --- THINKING MODE PARAMETERS ---
-    thinking_mode = data.get("thinking_mode", False) # True or False
-    thinking_effort = data.get("thinking_effort", "medium") # "low", "medium", "high"
+    # 🧠 --- THINKING MODE ---
+    thinking_mode = data.get("thinking_mode", False)
+    thinking_effort = data.get("thinking_effort", "medium")
 
-    # 🕒 --- REAL-TIME IST INJECTION ---
+    # 🕒 --- REAL-TIME IST ---
     ist_time = datetime.now(timezone.utc) + timedelta(hours=5, minutes=30)
     current_date = ist_time.strftime("%A, %d %B %Y, %I:%M %p IST")
 
@@ -90,70 +71,69 @@ def chat():
     thinking_instruction = ""
     if thinking_mode:
         if thinking_effort == "low":
-            thinking_instruction = "\n[SYSTEM: THINKING MODE ENABLED (LOW EFFORT) - Keep your internal reasoning brief, direct, and fast before answering.]"
+            thinking_instruction = "\n[SYSTEM: THINKING MODE (LOW) - Reason briefly and directly before answering.]"
         elif thinking_effort == "high":
-            thinking_instruction = "\n[SYSTEM: THINKING MODE ENABLED (HIGH EFFORT) - Perform a deep, exhaustive, and highly detailed step-by-step analysis. Consider edge cases, multiple perspectives, and logic verification before providing the final answer.]"
-        else: # medium
-            thinking_instruction = "\n[SYSTEM: THINKING MODE ENABLED (MEDIUM EFFORT) - Perform a standard step-by-step logical reasoning process before answering.]"
+            thinking_instruction = "\n[SYSTEM: THINKING MODE (HIGH) - Perform exhaustive, multi-step logical reasoning before answering.]"
+        else:
+            thinking_instruction = "\n[SYSTEM: THINKING MODE (MEDIUM) - Perform a standard step-by-step reasoning process.]"
 
     # 🧠 --- GOD MODE SYSTEM PROMPT ---
-    system_prompt = f"""
-    You are CODE VED, an advanced AI System engineered EXCLUSIVELY by DIVY PATEL.
-    Current Live Date and Time: {current_date}.
-    {thinking_instruction}
+    system_prompt = f"""You are CODE VED, an advanced AI System engineered EXCLUSIVELY by DIVY PATEL.
+    Current Live Date and Time: {current_date}.{thinking_instruction}
     
     STRICT DIRECTIVES:
     1. NEVER invent, guess, or hallucinate product launches, dates, news, or facts.
     2. If you receive "LIVE WEB SEARCH RESULTS", you MUST base your answer ENTIRELY on that data.
-    3. Do NOT say "Based on the provided search results". Just answer naturally and confidently, citing the sources/links if needed.
+    3. Do NOT say "Based on the provided search results". Just answer naturally and confidently.
     """
 
-    # 🚀 --- AUTO-SEARCH INJECTION (USING SERPAPI) ---
+    # 🚀 --- AUTO-SEARCH INJECTION ---
     if is_search:
         scraped_data = web_search_scraper(user_message)
-        
         search_context = "\n\n--- [LIVE VERIFIED GOOGLE SEARCH DATA] ---\n"
         if scraped_data:
             for idx, res in enumerate(scraped_data):
                 search_context += f"{idx+1}. TITLE: {res['title']}\nSNIPPET: {res['snippet']}\nURL: {res['link']}\n\n"
-            
-            search_context += "[SYSTEM COMMAND: Use the above live Google data to answer the user accurately. Synthesize the info naturally.]"
+            search_context += "[SYSTEM COMMAND: Use the above live Google data to answer the user accurately.]"
         else:
-            search_context += "[SYSTEM ALERT: Live search did not return results. Rely on your existing knowledge, but DO NOT hallucinate recent news.]"
+            search_context += "[SYSTEM ALERT: Live search did not return results. Rely on your existing knowledge, but DO NOT hallucinate.]"
             
         user_message = f"USER QUERY: {user_message}\n\n{search_context}"
 
     messages = [{"role": "system", "content": system_prompt}]
 
+    # ⚠️ FIX: Remove the last message from history because the frontend already adds the current query to it!
+    # Without this, the model gets the same question twice and gets confused.
+    if history and history[-1].get("role") == "user":
+        history = history[:-1]
+
     for msg in history:
         role = msg.get("role", "user")
         content = msg.get("content", "")
         if content:
             messages.append({"role": role, "content": content})
 
-    content_payload = []
-    if user_message.strip():
-        content_payload.append({"type": "text", "text": user_message})
-        
-    for att in attachments:
-        att_type = att.get("type")
-        b64_data = att.get("data")
-        if att_type == "image":
-            content_payload.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64_data}"}})
-        elif att_type in ["audio", "file"]:
-            content_payload.append({"type": "input_audio", "input_audio": {"data": b64_data, "format": "wav"}})
-
-    if not content_payload:
-        content_payload.append({"type": "text", "text": "Hello"})
-
-    messages.append({"role": "user", "content": content_payload})
+    # ⚠️ BIG MISTAKE FIXED: Only use array format if there are actually attachments.
+    # Otherwise, send pure string text so standard LLMs don't break their system prompt alignment!
+    if attachments:
+        content_payload = [{"type": "text", "text": user_message}]
+        for att in attachments:
+            att_type = att.get("type")
+            b64_data = att.get("data")
+            if att_type == "image":
+                content_payload.append({"type": "image_url", "image_url": {"url": f"data:image/jpeg;base64,{b64_data}"}})
+            elif att_type in ["audio", "file"]:
+                content_payload.append({"type": "input_audio", "input_audio": {"data": b64_data, "format": "wav"}})
+        messages.append({"role": "user", "content": content_payload})
+    else:
+        # Standard text models perfectly understand this
+        messages.append({"role": "user", "content": user_message})
 
     headers = {
         "Authorization": f"Bearer {API_KEY}", 
         "Accept": "text/event-stream"
     }
 
-    # 🌟 --- NVIDIA FORMAT INTEGRATED ---
     payload = {
         "model": MODEL_ID,
         "messages": messages,
@@ -161,7 +141,6 @@ def chat():
         "temperature": float(temperature),
         "top_p": 0.95,
         "stream": True,
-        # Enable native thinking in the backend if the user requested it
         "chat_template_kwargs": {"enable_thinking": thinking_mode}
     }