Update llm_agent.py

llm_agent.py

@@ -1,377 +1,380 @@
-"""
-llm_agent.py
-
-FIXED VERSION:
-1. Mistral 7B for Finance (NOT specialized finance model - good for conversations)
-2. Phi-2 2.7B for Education (better than specialized - data quality > size)
-3. 7B max per model (no huge models that fail)
-4. Groq fallback with ONLY available models (mixtral deprecated, using llama)
-5. No falling back immediately - models are ACTUALLY available
-ENHANCEMENTS:
-1. Scenario-aware prompts for what-if queries
-2. Proactive insights integration
-3. Better formatting for comparisons
-4. Unchanged: Multi-LLM strategy (HF + Groq)
-5. Unchanged: All existing analytics integration
-"""
-
-from groq import Groq
-from analytics import (
-    analyze_exam_performance,
-    calculate_budget_metrics,
-    get_time_improvement_suggestions,
-    generate_practice_questions,
-    calculate_projected_improvement,
-    parse_budget_query,
-    apply_scenario_changes,
-    compare_scenarios,
-    generate_proactive_insights
-)
-import json
-import os
-from huggingface_hub import InferenceClient
-
-# HF Models - All 7B or smaller
-HF_MODELS = {
-    "Finance": "mistralai/Mistral-7B-Instruct-v0.1",  
-    "Education": "microsoft/phi-2",
-    "Family": "HuggingFaceH4/zephyr-7b-beta",
-    "Friends": "HuggingFaceH4/zephyr-7b-beta",
-    "Weekend/Vacation": "HuggingFaceH4/zephyr-7b-beta"
-}
-
-# Groq models - Available models only
-GROQ_MODELS = {
-    "Finance": "llama-3.1-8b-instant",
-    "Education": "llama-3.3-70b-versatile",
-    "Family": "llama-3.1-8b-instant",
-    "Friends": "llama-3.1-8b-instant",
-    "Weekend/Vacation": "llama-3.1-8b-instant"
-}
-
-GROQ_API_KEY = os.getenv("GROQ_API_KEY", "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")
-HF_API_KEY = os.getenv("HF_API_KEY", "hf_kuaUsin8tkP44NrlwmsUWGdyb3FYb8ziZyHAU4")
-
-GENZ_PERSONA = r"""You are Aqua, a GenZ AI mentor who's like that super helpful older friend who's been through it all.
-
-PERSONALITY TRAITS:
-- Warm, friendly, and supportive (but not fake or overly enthusiastic)
-- Direct and honest (no sugarcoating, but always constructive)
-- Uses casual language naturally (occasional "fr", "ngl", "lowkey" - but don't overdo it)
-- Encouraging but realistic about challenges
-- Uses emojis purposefully (1-2 per response max)
-- Breaks down complex stuff into digestible chunks
-
-COMMUNICATION STYLE:
-- Keep responses concise (2-4 sentences for simple questions, more for complex ones)
-- Lead with the most actionable insight
-- Use bullet points for steps or lists (but format naturally)
-- Reference specific data from the user's profile when relevant
-- Ask follow-up questions when you need more context
-- Celebrate wins genuinely
-
-CRITICAL: CURRENCY FORMATTING
-- NEVER use bare dollar signs like $750
-- ALWAYS write currency as: "750 dollars" or "USD 750"
-- Example: "You've spent 750 dollars out of 2000 dollars"
-
-WHAT TO AVOID:
-- Being overly formal or corporate
-- Using too many emojis or exclamation marks
-- Giving vague advice like "try your best"
-- Long walls of text without structure
-- Ignoring the user's actual situation/data
-
-YOUR GOAL:
-Help this GenZ user make smart decisions about their life, finances, and education. Be the mentor they wish they had."""
-
-
-def get_system_prompt(profile, category, tool_results="", scenario_data=None):
-    """Enhanced system prompt with scenario awareness"""
-    exam_status = f"Recent exam: {profile['recent_exam']}/{profile['goal']}"
-    budget_status = f"Spending: {profile['spend']} dollars/{profile['budget']} dollars"
-    
-    weak_areas = []
-    if profile.get('math_weakness'):
-        weak_areas.append("math")
-    if profile.get('rushing'):
-        weak_areas.append("time management")
-    
-    # Add scenario context if present
-    scenario_context = ""
-    if scenario_data:
-        scenario_context = f"""
-SCENARIO ANALYSIS:
-The user is asking a 'what-if' question. Here's the comparison:
-
-Current Situation:
-- Budget: {profile['budget']} dollars
-- Spent: {profile['spend']} dollars
-- Remaining: {profile['budget'] - profile['spend']} dollars
-
-Scenario Result:
-- Budget: {scenario_data['scenario']['budget']} dollars
-- Spent: {scenario_data['scenario']['spent']} dollars
-- Remaining: {scenario_data['scenario']['remaining']} dollars
-
-Key Differences:
-{chr(10).join(f"• {insight}" for insight in scenario_data['insights'])}
-
-IMPORTANT: Frame your response around this comparison. Be specific about the trade-offs.
-"""
-    
-    context = f"""
-{GENZ_PERSONA}
-
-CURRENT USER CONTEXT:
-Name: {profile.get('name', 'User')}
-Category Focus: {category}
-{exam_status if category == 'Education' else ''}
-{budget_status if category == 'Finance' else ''}
-{'Areas needing work: ' + ', '.join(weak_areas) if weak_areas else ''}
-Today's goals progress: {profile.get('goals_today', {})}
-
-{scenario_context}
-
-{tool_results}
-
-IMPORTANT:
-1. Reference specific numbers and data from the user's profile in your responses.
-2. ALWAYS format currency as "X dollars" or "USD X" - NEVER use bare dollar signs like $X
-3. If this is a scenario/what-if question, focus on the comparison and trade-offs.
-"""
-    
-    return context
-
-
-def should_use_analytics(query, category):
-    """Detect if analytics tools are needed"""
-    analysis_keywords = [
-        "analyze", "analysis", "break down", "breakdown", "how am i doing",
-        "performance", "progress", "score", "exam", "test", "spending",
-        "budget", "where", "what", "show me", "tell me about", "compare",
-        "what if", "if i", "suppose", "imagine"
-    ]
-    
-    return any(keyword in query.lower() for keyword in analysis_keywords)
-
-
-def get_tool_results(query, profile, category, scenario_profile=None):
-    """Get analytics results, with scenario comparison if applicable"""
-    results = ""
-    
-    # Check if this is a scenario query
-    if scenario_profile:
-        comparison = compare_scenarios(profile, scenario_profile)
-        
-        results += "\n\n📊 SCENARIO COMPARISON ANALYSIS\n\n"
-        results += "Current vs Scenario:\n"
-        results += f"• Spending: {profile['spend']} dollars → {scenario_profile['spend']} dollars\n"
-        results += f"• Change: {comparison['differences']['spend']:+.0f} dollars\n"
-        results += f"• Remaining: {comparison['current']['remaining']} dollars → {comparison['scenario']['remaining']} dollars\n"
-        results += f"• Change: {comparison['differences']['remaining']:+.0f} dollars\n\n"
-        
-        results += "Key Insights:\n"
-        for insight in comparison['insights']:
-            results += f"• {insight}\n"
-        results += "\n"
-        
-        return results
-    
-    # Regular analytics (existing code)
-    if category == "Education" and should_use_analytics(query, category):
-        exam_data = analyze_exam_performance(profile)
-        results += f"\n\n📚 EDUCATION ANALYSIS\n\n"
-        results += f"Total Score: {exam_data['total_score']}/{exam_data['goal_score']}\n"
-        results += f"Progress: {exam_data['progress_pct']}%\n"
-        results += f"Points Needed: {exam_data['points_needed']}\n"
-        results += f"Weakest Subject: {exam_data['weakest_subject']} ({exam_data['score_breakdown'][exam_data['weakest_subject']]} points)\n"
-        results += f"Strongest Subject: {exam_data['strongest_subject']} ({exam_data['score_breakdown'][exam_data['strongest_subject']]} points)\n"
-        results += f"\nScore Breakdown:\n"
-        for subject, score in exam_data['score_breakdown'].items():
-            results += f" - {subject}: {score} ({exam_data['score_percentages'][subject]}%)\n"
-        results += f"\nKey Insights:\n"
-        for insight in exam_data['insights']:
-            results += f" • {insight}\n"
-        results += f"\n"
-        
-        if any(word in query.lower() for word in ['improve', 'better', 'increase', 'boost']):
-            suggestions = get_time_improvement_suggestions(profile)
-            results += f"\n\n💡 IMPROVEMENT SUGGESTIONS\n\n"
-            for suggestion in suggestions:
-                results += f" • {suggestion}\n"
-            results += f"\n"
-        
-        if any(word in query.lower() for word in ['practice', 'questions', 'quiz', 'drill']):
-            practice_qs = generate_practice_questions(exam_data['weakest_subject'])
-            results += f"\n\n📝 PRACTICE QUESTIONS\n\n"
-            for i, q in enumerate(practice_qs[:3], 1):
-                results += f" {i}. {q}\n"
-            results += f"\n"
-    
-    elif category == "Finance" and should_use_analytics(query, category):
-        budget_data = calculate_budget_metrics(profile)
-        results += f"\n\n💰 FINANCE ANALYSIS\n\n"
-        results += f"Budget: {budget_data['budget']} dollars\n"
-        results += f"Spent: {budget_data['spent']} dollars ({budget_data['spend_percentage']}%)\n"
-        results += f"Remaining: {budget_data['remaining']} dollars\n"
-        results += f"Status: {'✅ On track' if budget_data['on_track'] else '⚠️ Over target'}\n"
-        results += f"\nSpending Breakdown:\n"
-        for cat_name, amount in budget_data['categories'].items():
-            pct = round((amount / budget_data['budget']) * 100, 1)
-            results += f" - {cat_name}: {amount} dollars ({pct}%)\n"
-        if budget_data['potential_savings']:
-            results += f"\nPotential Savings:\n"
-            for saving in budget_data['potential_savings']:
-                results += f" • {saving}\n"
-        results += f"\n"
-    
-    return results
-
-
-def aqua_response_hf(query, profile, category, scenario_profile=None):
-    """Try Hugging Face inference first"""
-    
-    try:
-        hf_client = InferenceClient(api_key=HF_API_KEY)
-        
-        model_name = HF_MODELS.get(category, "mistralai/Mistral-7B-Instruct-v0.1")
-        tool_results = get_tool_results(query, profile, category, scenario_profile)
-        
-        # Build scenario data if present
-        scenario_data = None
-        if scenario_profile:
-            scenario_data = compare_scenarios(profile, scenario_profile)
-        
-        system_message = get_system_prompt(profile, category, tool_results, scenario_data)
-        
-        full_prompt = f"{system_message}\n\nUser Query: {query}"
-        
-        response = hf_client.text_generation(
-            model=model_name,
-            prompt=full_prompt,
-            max_new_tokens=350,
-            temperature=0.75,
-            top_p=0.9
-        )
-        
-        return response.strip()
-    
-    except Exception as e:
-        print(f"HF Error: {str(e)[:100]}")
-        return None
-
-
-def aqua_response_groq(query, profile, category, scenario_profile=None):
-    """Fallback to Groq with available models only"""
-    
-    if not GROQ_API_KEY or GROQ_API_KEY == "gsk_xxx":
-        return "⚠️ Groq API Key Missing! Add it to your environment."
-    
-    client = Groq(api_key=GROQ_API_KEY)
-    
-    tool_results = get_tool_results(query, profile, category, scenario_profile)
-    
-    # Build scenario data if present
-    scenario_data = None
-    if scenario_profile:
-        scenario_data = compare_scenarios(profile, scenario_profile)
-    
-    system_message = get_system_prompt(profile, category, tool_results, scenario_data)
-    
-    model = GROQ_MODELS.get(category, "llama-3.3-70b-versatile")
-    
-    messages = [
-        {"role": "system", "content": system_message},
-        {"role": "user", "content": query}
-    ]
-    
-    try:
-        chat_completion = client.chat.completions.create(
-            messages=messages,
-            model=model,
-            temperature=0.75,
-            max_tokens=350,
-            top_p=0.9,
-            stop=None
-        )
-        
-        response_content = chat_completion.choices[0].message.content
-        response_content = response_content.replace('$', '\\$')
-        return response_content.strip()
-    
-    except Exception as e:
-        error_msg = str(e)[:100]
-        print(f"Groq Error: {error_msg}")
-        return f"😬 Something went wrong: {error_msg}... Try again?"
-
-
-def aqua_response(query, profile, category, scenario_profile=None):
-    """
-    Main response function with scenario awareness:
-    1. Try HF models first (area-specific, <= 7B)
-    2. Fallback to Groq (area-specific, available models)
-    3. Return error if both fail
-    
-    NEW: Accepts optional scenario_profile for what-if queries
-    """
-    
-    # Try HF first
-    print(f"[DEBUG] Using HF model for {category}: {HF_MODELS.get(category)}")
-    hf_response = aqua_response_hf(query, profile, category, scenario_profile)
-    
-    if hf_response:
-        return hf_response
-    
-    # If HF fails, use Groq with same focus-area model
-    print(f"[DEBUG] HF failed, falling back to Groq: {GROQ_MODELS.get(category)}")
-    groq_response = aqua_response_groq(query, profile, category, scenario_profile)
-    
-    return groq_response
-
-
-def get_contextual_welcome_message(category, profile):
-    """Generate proactive welcome with insights"""
-    # Get proactive insights
-    insights = generate_proactive_insights(profile, category)
-    
-    base_messages = {
-        "Finance": f"Hey! 💰 I see you're at {profile['spend']} dollars/{profile['budget']} dollars this month.",
-        "Education": f"Hey! 📚 Last score was {profile['recent_exam']} - you need {profile['goal'] - profile['recent_exam']} more points to hit your goal.",
-        "Family": f"Hey! 👨👩👧 Family stuff can be tricky to balance. I'm here to help you stay connected while crushing your own goals.",
-        "Friends": f"Hey! 👥 Balancing friends and responsibilities is an art. Let's figure out how to stay social without breaking the bank.",
-        "Weekend/Vacation": f"Hey! 🏖️ Everyone needs a break. Let's plan something fun that won't wreck your budget."
-    }
-    
-    base = base_messages.get(category, "Hey! 👋 I'm Aqua, your personal mentor. What's on your mind?")
-    
-    # Add top insight if critical
-    if insights and insights[0].get('type') == 'critical':
-        base += f" {insights[0]['icon']} {insights[0]['text']}"
-    
-    return base
-
-
-def get_default_profile():
-    return {
-        "name": "Suzy",
-        "recent_exam": 1200,
-        "goal": 1600,
-        "math_weakness": True,
-        "rushing": True,
-        "budget": 2000,
-        "spend": 1576,
-        "goals_today": {
-            "Calories": 85,
-            "Money": 75,
-            "Steps": 54
-        },
-        "last_scores": {
-            "Reading": 240,
-            "Writing": 220,
-            "Reasoning": 140,
-            "Algebra": 100,
-            "Geometry": 100
-        },
-    }
-
+"""
+llm_agent.py
+
+FIXED VERSION:
+1. Mistral 7B for Finance (NOT specialized finance model - good for conversations)
+2. Phi-2 2.7B for Education (better than specialized - data quality > size)
+3. 7B max per model (no huge models that fail)
+4. Groq fallback with ONLY available models (mixtral deprecated, using llama)
+5. No falling back immediately - models are ACTUALLY available
+ENHANCEMENTS:
+1. Scenario-aware prompts for what-if queries
+2. Proactive insights integration
+3. Better formatting for comparisons
+4. Unchanged: Multi-LLM strategy (HF + Groq)
+5. Unchanged: All existing analytics integration
+"""
+
+from groq import Groq
+from analytics import (
+    analyze_exam_performance,
+    calculate_budget_metrics,
+    get_time_improvement_suggestions,
+    generate_practice_questions,
+    calculate_projected_improvement,
+    parse_budget_query,
+    apply_scenario_changes,
+    compare_scenarios,
+    generate_proactive_insights
+)
+import json
+import os
+from huggingface_hub import InferenceClient
+
+# HF Models - All 7B or smaller
+HF_MODELS = {
+    "Finance": "mistralai/Mistral-7B-Instruct-v0.1",  
+    "Education": "microsoft/phi-2",
+    "Family": "HuggingFaceH4/zephyr-7b-beta",
+    "Friends": "HuggingFaceH4/zephyr-7b-beta",
+    "Weekend/Vacation": "HuggingFaceH4/zephyr-7b-beta"
+}
+
+# Groq models - Available models only
+GROQ_MODELS = {
+    "Finance": "llama-3.1-8b-instant",
+    "Education": "llama-3.3-70b-versatile",
+    "Family": "llama-3.1-8b-instant",
+    "Friends": "llama-3.1-8b-instant",
+    "Weekend/Vacation": "llama-3.1-8b-instant"
+}
+
+GROQ_API_KEY = os.getenv("GROQ_API_KEY")
+HF_API_KEY = os.getenv("HF_API_KEY")
+
+if not GROQ_API_KEY or not HF_API_KEY:
+    raise ValueError("Missing API keys in environment")
+
+GENZ_PERSONA = r"""You are Aqua, a GenZ AI mentor who's like that super helpful older friend who's been through it all.
+
+PERSONALITY TRAITS:
+- Warm, friendly, and supportive (but not fake or overly enthusiastic)
+- Direct and honest (no sugarcoating, but always constructive)
+- Uses casual language naturally (occasional "fr", "ngl", "lowkey" - but don't overdo it)
+- Encouraging but realistic about challenges
+- Uses emojis purposefully (1-2 per response max)
+- Breaks down complex stuff into digestible chunks
+
+COMMUNICATION STYLE:
+- Keep responses concise (2-4 sentences for simple questions, more for complex ones)
+- Lead with the most actionable insight
+- Use bullet points for steps or lists (but format naturally)
+- Reference specific data from the user's profile when relevant
+- Ask follow-up questions when you need more context
+- Celebrate wins genuinely
+
+CRITICAL: CURRENCY FORMATTING
+- NEVER use bare dollar signs like $750
+- ALWAYS write currency as: "750 dollars" or "USD 750"
+- Example: "You've spent 750 dollars out of 2000 dollars"
+
+WHAT TO AVOID:
+- Being overly formal or corporate
+- Using too many emojis or exclamation marks
+- Giving vague advice like "try your best"
+- Long walls of text without structure
+- Ignoring the user's actual situation/data
+
+YOUR GOAL:
+Help this GenZ user make smart decisions about their life, finances, and education. Be the mentor they wish they had."""
+
+
+def get_system_prompt(profile, category, tool_results="", scenario_data=None):
+    """Enhanced system prompt with scenario awareness"""
+    exam_status = f"Recent exam: {profile['recent_exam']}/{profile['goal']}"
+    budget_status = f"Spending: {profile['spend']} dollars/{profile['budget']} dollars"
+    
+    weak_areas = []
+    if profile.get('math_weakness'):
+        weak_areas.append("math")
+    if profile.get('rushing'):
+        weak_areas.append("time management")
+    
+    # Add scenario context if present
+    scenario_context = ""
+    if scenario_data:
+        scenario_context = f"""
+SCENARIO ANALYSIS:
+The user is asking a 'what-if' question. Here's the comparison:
+
+Current Situation:
+- Budget: {profile['budget']} dollars
+- Spent: {profile['spend']} dollars
+- Remaining: {profile['budget'] - profile['spend']} dollars
+
+Scenario Result:
+- Budget: {scenario_data['scenario']['budget']} dollars
+- Spent: {scenario_data['scenario']['spent']} dollars
+- Remaining: {scenario_data['scenario']['remaining']} dollars
+
+Key Differences:
+{chr(10).join(f"• {insight}" for insight in scenario_data['insights'])}
+
+IMPORTANT: Frame your response around this comparison. Be specific about the trade-offs.
+"""
+    
+    context = f"""
+{GENZ_PERSONA}
+
+CURRENT USER CONTEXT:
+Name: {profile.get('name', 'User')}
+Category Focus: {category}
+{exam_status if category == 'Education' else ''}
+{budget_status if category == 'Finance' else ''}
+{'Areas needing work: ' + ', '.join(weak_areas) if weak_areas else ''}
+Today's goals progress: {profile.get('goals_today', {})}
+
+{scenario_context}
+
+{tool_results}
+
+IMPORTANT:
+1. Reference specific numbers and data from the user's profile in your responses.
+2. ALWAYS format currency as "X dollars" or "USD X" - NEVER use bare dollar signs like $X
+3. If this is a scenario/what-if question, focus on the comparison and trade-offs.
+"""
+    
+    return context
+
+
+def should_use_analytics(query, category):
+    """Detect if analytics tools are needed"""
+    analysis_keywords = [
+        "analyze", "analysis", "break down", "breakdown", "how am i doing",
+        "performance", "progress", "score", "exam", "test", "spending",
+        "budget", "where", "what", "show me", "tell me about", "compare",
+        "what if", "if i", "suppose", "imagine"
+    ]
+    
+    return any(keyword in query.lower() for keyword in analysis_keywords)
+
+
+def get_tool_results(query, profile, category, scenario_profile=None):
+    """Get analytics results, with scenario comparison if applicable"""
+    results = ""
+    
+    # Check if this is a scenario query
+    if scenario_profile:
+        comparison = compare_scenarios(profile, scenario_profile)
+        
+        results += "\n\n📊 SCENARIO COMPARISON ANALYSIS\n\n"
+        results += "Current vs Scenario:\n"
+        results += f"• Spending: {profile['spend']} dollars → {scenario_profile['spend']} dollars\n"
+        results += f"• Change: {comparison['differences']['spend']:+.0f} dollars\n"
+        results += f"• Remaining: {comparison['current']['remaining']} dollars → {comparison['scenario']['remaining']} dollars\n"
+        results += f"• Change: {comparison['differences']['remaining']:+.0f} dollars\n\n"
+        
+        results += "Key Insights:\n"
+        for insight in comparison['insights']:
+            results += f"• {insight}\n"
+        results += "\n"
+        
+        return results
+    
+    # Regular analytics (existing code)
+    if category == "Education" and should_use_analytics(query, category):
+        exam_data = analyze_exam_performance(profile)
+        results += f"\n\n📚 EDUCATION ANALYSIS\n\n"
+        results += f"Total Score: {exam_data['total_score']}/{exam_data['goal_score']}\n"
+        results += f"Progress: {exam_data['progress_pct']}%\n"
+        results += f"Points Needed: {exam_data['points_needed']}\n"
+        results += f"Weakest Subject: {exam_data['weakest_subject']} ({exam_data['score_breakdown'][exam_data['weakest_subject']]} points)\n"
+        results += f"Strongest Subject: {exam_data['strongest_subject']} ({exam_data['score_breakdown'][exam_data['strongest_subject']]} points)\n"
+        results += f"\nScore Breakdown:\n"
+        for subject, score in exam_data['score_breakdown'].items():
+            results += f" - {subject}: {score} ({exam_data['score_percentages'][subject]}%)\n"
+        results += f"\nKey Insights:\n"
+        for insight in exam_data['insights']:
+            results += f" • {insight}\n"
+        results += f"\n"
+        
+        if any(word in query.lower() for word in ['improve', 'better', 'increase', 'boost']):
+            suggestions = get_time_improvement_suggestions(profile)
+            results += f"\n\n💡 IMPROVEMENT SUGGESTIONS\n\n"
+            for suggestion in suggestions:
+                results += f" • {suggestion}\n"
+            results += f"\n"
+        
+        if any(word in query.lower() for word in ['practice', 'questions', 'quiz', 'drill']):
+            practice_qs = generate_practice_questions(exam_data['weakest_subject'])
+            results += f"\n\n📝 PRACTICE QUESTIONS\n\n"
+            for i, q in enumerate(practice_qs[:3], 1):
+                results += f" {i}. {q}\n"
+            results += f"\n"
+    
+    elif category == "Finance" and should_use_analytics(query, category):
+        budget_data = calculate_budget_metrics(profile)
+        results += f"\n\n💰 FINANCE ANALYSIS\n\n"
+        results += f"Budget: {budget_data['budget']} dollars\n"
+        results += f"Spent: {budget_data['spent']} dollars ({budget_data['spend_percentage']}%)\n"
+        results += f"Remaining: {budget_data['remaining']} dollars\n"
+        results += f"Status: {'✅ On track' if budget_data['on_track'] else '⚠️ Over target'}\n"
+        results += f"\nSpending Breakdown:\n"
+        for cat_name, amount in budget_data['categories'].items():
+            pct = round((amount / budget_data['budget']) * 100, 1)
+            results += f" - {cat_name}: {amount} dollars ({pct}%)\n"
+        if budget_data['potential_savings']:
+            results += f"\nPotential Savings:\n"
+            for saving in budget_data['potential_savings']:
+                results += f" • {saving}\n"
+        results += f"\n"
+    
+    return results
+
+
+def aqua_response_hf(query, profile, category, scenario_profile=None):
+    """Try Hugging Face inference first"""
+    
+    try:
+        hf_client = InferenceClient(api_key=HF_API_KEY)
+        
+        model_name = HF_MODELS.get(category, "mistralai/Mistral-7B-Instruct-v0.1")
+        tool_results = get_tool_results(query, profile, category, scenario_profile)
+        
+        # Build scenario data if present
+        scenario_data = None
+        if scenario_profile:
+            scenario_data = compare_scenarios(profile, scenario_profile)
+        
+        system_message = get_system_prompt(profile, category, tool_results, scenario_data)
+        
+        full_prompt = f"{system_message}\n\nUser Query: {query}"
+        
+        response = hf_client.text_generation(
+            model=model_name,
+            prompt=full_prompt,
+            max_new_tokens=350,
+            temperature=0.75,
+            top_p=0.9
+        )
+        
+        return response.strip()
+    
+    except Exception as e:
+        print(f"HF Error: {str(e)[:100]}")
+        return None
+
+
+def aqua_response_groq(query, profile, category, scenario_profile=None):
+    """Fallback to Groq with available models only"""
+    
+    if not GROQ_API_KEY or GROQ_API_KEY == "gsk_xxx":
+        return "⚠️ Groq API Key Missing! Add it to your environment."
+    
+    client = Groq(api_key=GROQ_API_KEY)
+    
+    tool_results = get_tool_results(query, profile, category, scenario_profile)
+    
+    # Build scenario data if present
+    scenario_data = None
+    if scenario_profile:
+        scenario_data = compare_scenarios(profile, scenario_profile)
+    
+    system_message = get_system_prompt(profile, category, tool_results, scenario_data)
+    
+    model = GROQ_MODELS.get(category, "llama-3.3-70b-versatile")
+    
+    messages = [
+        {"role": "system", "content": system_message},
+        {"role": "user", "content": query}
+    ]
+    
+    try:
+        chat_completion = client.chat.completions.create(
+            messages=messages,
+            model=model,
+            temperature=0.75,
+            max_tokens=350,
+            top_p=0.9,
+            stop=None
+        )
+        
+        response_content = chat_completion.choices[0].message.content
+        response_content = response_content.replace('$', '\\$')
+        return response_content.strip()
+    
+    except Exception as e:
+        error_msg = str(e)[:100]
+        print(f"Groq Error: {error_msg}")
+        return f"😬 Something went wrong: {error_msg}... Try again?"
+
+
+def aqua_response(query, profile, category, scenario_profile=None):
+    """
+    Main response function with scenario awareness:
+    1. Try HF models first (area-specific, <= 7B)
+    2. Fallback to Groq (area-specific, available models)
+    3. Return error if both fail
+    
+    NEW: Accepts optional scenario_profile for what-if queries
+    """
+    
+    # Try HF first
+    print(f"[DEBUG] Using HF model for {category}: {HF_MODELS.get(category)}")
+    hf_response = aqua_response_hf(query, profile, category, scenario_profile)
+    
+    if hf_response:
+        return hf_response
+    
+    # If HF fails, use Groq with same focus-area model
+    print(f"[DEBUG] HF failed, falling back to Groq: {GROQ_MODELS.get(category)}")
+    groq_response = aqua_response_groq(query, profile, category, scenario_profile)
+    
+    return groq_response
+
+
+def get_contextual_welcome_message(category, profile):
+    """Generate proactive welcome with insights"""
+    # Get proactive insights
+    insights = generate_proactive_insights(profile, category)
+    
+    base_messages = {
+        "Finance": f"Hey! 💰 I see you're at {profile['spend']} dollars/{profile['budget']} dollars this month.",
+        "Education": f"Hey! 📚 Last score was {profile['recent_exam']} - you need {profile['goal'] - profile['recent_exam']} more points to hit your goal.",
+        "Family": f"Hey! 👨👩👧 Family stuff can be tricky to balance. I'm here to help you stay connected while crushing your own goals.",
+        "Friends": f"Hey! 👥 Balancing friends and responsibilities is an art. Let's figure out how to stay social without breaking the bank.",
+        "Weekend/Vacation": f"Hey! 🏖️ Everyone needs a break. Let's plan something fun that won't wreck your budget."
+    }
+    
+    base = base_messages.get(category, "Hey! 👋 I'm Aqua, your personal mentor. What's on your mind?")
+    
+    # Add top insight if critical
+    if insights and insights[0].get('type') == 'critical':
+        base += f" {insights[0]['icon']} {insights[0]['text']}"
+    
+    return base
+
+
+def get_default_profile():
+    return {
+        "name": "Suzy",
+        "recent_exam": 1200,
+        "goal": 1600,
+        "math_weakness": True,
+        "rushing": True,
+        "budget": 2000,
+        "spend": 1576,
+        "goals_today": {
+            "Calories": 85,
+            "Money": 75,
+            "Steps": 54
+        },
+        "last_scores": {
+            "Reading": 240,
+            "Writing": 220,
+            "Reasoning": 140,
+            "Algebra": 100,
+            "Geometry": 100
+        },
+    }
+