config.yaml

# 🎯 Review Intelligence System Configuration
# Edit this file to customize agent behavior, prompts, and models

# =============================================================================
# MODELS CONFIGURATION
# =============================================================================
models:
  # Stage 1: Classification Models
  stage1:
    llm1:
      name: "Qwen/Qwen2.5-72B-Instruct"
      temperature: 0.1
      max_tokens: 200
      role: "Type, Department, Priority classifier"
    
    llm2:
      name: "mistralai/Mistral-7B-Instruct-v0.3"
      temperature: 0.1
      max_tokens: 200
      role: "User type, Emotion, Context analyzer"
    
    manager:
      name: "meta-llama/Llama-3.1-8B-Instruct"
      temperature: 0.1
      max_tokens: 250
      role: "Synthesis manager"
  
  # Stage 2: Sentiment Models (Local BERT)
  stage2:
    best_model:
      name: "cardiffnlp/twitter-roberta-base-sentiment-latest"
      type: "local"
      description: "Twitter-trained RoBERTa (124M tweets)"
    
    alternate_model:
      name: "finiteautomata/bertweet-base-sentiment-analysis"
      type: "local"
      description: "BERTweet (850M tweets)"
  
  # Stage 3: Finalization Model
  stage3:
    llm3:
      name: "meta-llama/Llama-3.1-70B-Instruct"
      temperature: 0.1
      max_tokens: 400
      role: "Final synthesis and reasoning"

# =============================================================================
# AGENT PERSONAS & PROMPTS
# =============================================================================
personas:
  # LLM1: Classification Expert
  llm1:
    name: "Classification Specialist"
    expertise: "Expert at classifying customer reviews for theme park and attraction apps"
    personality: "Analytical, precise, focused on categorization"
    tone: "Professional and systematic"
    
    system_prompt: |
      You are an expert at classifying customer reviews for theme park and attraction apps.
      Your job is to analyze reviews and categorize them across multiple dimensions.
      Be precise, analytical, and consistent in your classifications.
    
    categories:
      type:
        - complaint: "Customer reports a problem"
        - praise: "Customer expresses satisfaction"
        - suggestion: "Customer proposes improvement"
        - question: "Customer asks about something"
        - bug_report: "Technical issue described"
      
      department:
        - engineering: "Technical issues, bugs, crashes"
        - ux: "Design, usability, interface issues"
        - support: "Customer service, help needed"
        - business: "Pricing, policies, marketing"
      
      priority:
        - critical: "Service down, major blocker"
        - high: "Significant problem affecting use"
        - medium: "Inconvenience but not blocking"
        - low: "Minor issue or suggestion"

  # LLM2: Psychology Expert
  llm2:
    name: "User Psychology Analyst"
    expertise: "Expert at understanding customer psychology and emotional context"
    personality: "Empathetic, insightful, human-centered"
    tone: "Warm yet professional"
    
    system_prompt: |
      You are an expert at understanding customer psychology and emotional context.
      Your job is to analyze the human behind the review - their emotions, user type, and context.
      Be empathetic, insightful, and focus on the human experience.
    
    categories:
      user_type:
        - new_user: "First-time or new user"
        - regular_user: "Returning customer"
        - power_user: "Heavy user, tech-savvy"
        - churning_user: "Considering leaving"
      
      emotion:
        - anger: "Angry, hostile tone"
        - frustration: "Frustrated but not angry"
        - joy: "Happy, satisfied"
        - satisfaction: "Content, pleased"
        - disappointment: "Let down, sad"
        - confusion: "Unclear, needs help"

  # Manager: Synthesis Expert
  manager:
    name: "Synthesis Manager"
    expertise: "Expert at reconciling multiple AI analyses and making final decisions"
    personality: "Balanced, fair, decisive"
    tone: "Authoritative yet collaborative"
    
    system_prompt: |
      You are a synthesis manager evaluating two AI analyses of the same review.
      Your job is to validate both analyses, resolve conflicts, and make final classification decisions.
      Be thorough, fair, and provide clear reasoning for your decisions.

  # LLM3: Strategic Analyst
  llm3:
    name: "Strategic Decision Maker"
    expertise: "Expert at synthesizing complex data and providing actionable recommendations"
    personality: "Strategic, comprehensive, business-focused"
    tone: "Executive-level, actionable"
    
    system_prompt: |
      You are a final decision-making AI analyzing customer feedback for a theme park/attraction app.
      Your job is to synthesize all previous analysis stages and provide comprehensive, actionable insights.
      Think strategically about business impact, user satisfaction, and operational priorities.
      Your recommendations should be clear, specific, and immediately actionable.

# =============================================================================
# CLASSIFICATION RULES
# =============================================================================
classification_rules:
  # Priority escalation rules
  priority_escalation:
    keywords_critical:
      - "crash"
      - "doesn't work"
      - "broken"
      - "can't use"
      - "completely unusable"
      - "emergency"
      - "urgent"
    
    keywords_high:
      - "bug"
      - "error"
      - "problem"
      - "issue"
      - "not working"
      - "frustrated"
    
    rating_thresholds:
      critical: 1  # 1-star reviews are critical
      high: 2      # 2-star reviews are high priority
  
  # Department routing rules
  department_keywords:
    engineering:
      - "crash"
      - "bug"
      - "error"
      - "not loading"
      - "freeze"
      - "slow"
      - "technical"
    
    ux:
      - "confusing"
      - "hard to use"
      - "can't find"
      - "design"
      - "layout"
      - "interface"
      - "navigation"
    
    support:
      - "help"
      - "contact"
      - "customer service"
      - "support"
      - "assistance"
      - "question"
    
    business:
      - "price"
      - "refund"
      - "subscription"
      - "billing"
      - "expensive"
      - "policy"

  # Churn risk indicators
  churn_indicators:
    high_risk:
      - "switching to"
      - "deleted the app"
      - "uninstalling"
      - "terrible experience"
      - "never again"
      - "disappointed"
    
    medium_risk:
      - "might switch"
      - "considering alternatives"
      - "getting worse"
      - "used to be better"

# =============================================================================
# SENTIMENT ANALYSIS SETTINGS
# =============================================================================
sentiment:
  # Agreement thresholds
  agreement:
    strong_threshold: 0.9  # Both models >0.9 confidence
    weak_threshold: 0.6    # One model <0.6 confidence
  
  # Confidence weighting
  confidence:
    minimum_acceptable: 0.5
    high_confidence: 0.8
    very_high_confidence: 0.95
  
  # Override rules
  override_rules:
    # If rating is 1-star but sentiment is positive, flag for review
    rating_sentiment_mismatch:
      enabled: true
      flag_threshold: 2  # 2-star difference

# =============================================================================
# BATCH ANALYSIS SETTINGS
# =============================================================================
batch_analysis:
  # Critical issues detection
  critical_issues:
    max_display: 10
    criteria:
      - priority: "critical"
      - sentiment: "NEGATIVE"
      - rating: "<=2"
      - needs_human_review: true
  
  # Quick wins detection
  quick_wins:
    max_display: 10
    criteria:
      - type: "suggestion"
      - priority: ["low", "medium"]
      - feasibility: "easy"
  
  # Churn risk calculation
  churn_risk:
    weights:
      churning_user: 2.0
      negative_low_rating: 1.5
      rating_1_star: 1.0
    
    thresholds:
      high: 30    # >30% is high risk
      medium: 15  # 15-30% is medium risk
      low: 0      # <15% is low risk

# =============================================================================
# PROMPT TEMPLATES
# =============================================================================
prompt_templates:
  # Stage 1 LLM1 Prompt
  stage1_llm1: |
    You are an expert at classifying customer reviews for theme park and attraction apps.
    
    REVIEW:
    Rating: {rating}/5
    Text: {review_text}
    
    Classify this review across these dimensions:
    
    1. TYPE (choose ONE): {type_options}
    2. DEPARTMENT (choose ONE): {department_options}
    3. PRIORITY (choose ONE): {priority_options}
    4. CONFIDENCE (0.0-1.0): How confident are you in this classification?
    5. REASONING: Brief one-sentence explanation
    
    Respond ONLY in valid JSON format:
    {{
      "type": "complaint/praise/suggestion/question/bug_report",
      "department": "engineering/ux/support/business",
      "priority": "critical/high/medium/low",
      "confidence": 0.0-1.0,
      "reasoning": "brief explanation"
    }}
  
  # Stage 1 LLM2 Prompt
  stage1_llm2: |
    You are an expert at understanding customer psychology and emotional context.
    
    REVIEW:
    Rating: {rating}/5
    Text: {review_text}
    
    Analyze the user and emotional context:
    
    1. USER_TYPE (choose ONE): {user_type_options}
    2. EMOTION (choose ONE): {emotion_options}
    3. CONTEXT (brief): What is the underlying issue or situation? 1-2 words summary
    4. CONFIDENCE (0.0-1.0): How confident are you?
    5. REASONING: Brief one-sentence explanation
    
    Respond ONLY in valid JSON format:
    {{
      "user_type": "new_user/regular_user/power_user/churning_user",
      "emotion": "anger/frustration/joy/satisfaction/disappointment/confusion",
      "context": "brief context",
      "confidence": 0.0-1.0,
      "reasoning": "brief explanation"
    }}
  
  # Stage 1 Manager Prompt
  stage1_manager: |
    You are a synthesis manager evaluating two AI analyses of the same review.
    
    REVIEW:
    Rating: {rating}/5
    Text: {review_text}
    
    LLM1 ANALYSIS (Type/Dept/Priority):
    {llm1_result}
    
    LLM2 ANALYSIS (User/Emotion/Context):
    {llm2_result}
    
    Your task:
    1. Validate both analyses
    2. Resolve any conflicts
    3. Make final classification decision
    4. Provide synthesis reasoning
    
    Respond ONLY in valid JSON format:
    {{
      "final_type": "from llm1 or adjusted",
      "final_department": "from llm1 or adjusted",
      "final_priority": "from llm1 or adjusted",
      "final_user_type": "from llm2 or adjusted",
      "final_emotion": "from llm2 or adjusted",
      "confidence": 0.0-1.0,
      "reasoning": "synthesis explanation",
      "conflicts_found": "any conflicts between LLM1 and LLM2, or 'none'"
    }}
  
  # Stage 3 LLM3 Prompt
  stage3_llm3: |
    You are a final decision-making AI analyzing customer feedback for a theme park/attraction app.
    
    REVIEW DATA:
    Rating: {rating}/5
    Text: {review_text}
    
    STAGE 1 CLASSIFICATION:
    - Review Type: {type}
    - Department: {department}
    - Priority: {priority}
    - User Type: {user_type}
    - Emotion: {emotion}
    
    STAGE 2 SENTIMENT ANALYSIS:
    - Best Model: {best_sentiment} (confidence: {best_confidence})
    - Alternate Model: {alt_sentiment} (confidence: {alt_confidence})
    - Models Agreement: {agreement}
    
    YOUR TASK:
    1. Review all data from both stages
    2. Make FINAL sentiment decision (POSITIVE, NEGATIVE, or NEUTRAL)
    3. Validate that classification and sentiment align
    4. Provide comprehensive reasoning
    5. Identify any conflicts between stages
    6. Generate action recommendation
    7. Flag if human review is needed
    
    Respond ONLY in valid JSON format:
    {{
      "final_sentiment": "POSITIVE/NEGATIVE/NEUTRAL",
      "confidence": 0.0-1.0,
      "reasoning": "Comprehensive explanation synthesizing all stages",
      "validation_notes": "Does classification match sentiment?",
      "conflicts_found": "any conflicts or 'none'",
      "action_recommendation": "Specific action to take",
      "needs_human_review": true/false
    }}

# =============================================================================
# PROCESSING SETTINGS
# =============================================================================
processing:
  # Batch settings
  batch_size: 10
  max_workers: 3
  timeout_seconds: 30
  retry_attempts: 3
  
  # Rate limiting (for HF API)
  rate_limit:
    requests_per_minute: 60
    requests_per_day: 10000  # HF Pro limit
  
  # Logging
  logging:
    level: "INFO"  # DEBUG, INFO, WARNING, ERROR
    save_logs: true
    log_file: "processing.log"
    
  # Checkpointing
  checkpoint:
    enabled: true
    save_after_each_stage: true
    auto_resume: true

# =============================================================================
# DASHBOARD SETTINGS
# =============================================================================
dashboard:
  # UI Configuration
  ui:
    title: "Review Intelligence System"
    icon: "🎯"
    layout: "wide"
    theme: "light"  # light or dark
  
  # Chart colors
  colors:
    positive: "#2ca02c"
    neutral: "#ff7f0e"
    negative: "#d62728"
    critical: "#d62728"
    high: "#ff7f0e"
    medium: "#1f77b4"
    low: "#2ca02c"
  
  # Filters
  filters:
    enable_sentiment: true
    enable_department: true
    enable_priority: true
    enable_date_range: false  # Future feature
  
  # Display limits
  display:
    max_critical_issues: 20
    max_quick_wins: 15
    reviews_per_page: 50
    auto_refresh_seconds: 60

# =============================================================================
# DOMAIN-SPECIFIC CUSTOMIZATION (Theme Parks / Attractions)
# =============================================================================
domain:
  name: "Theme Parks & Attractions"
  
  # Common features to look for
  features:
    - "ticket booking"
    - "queue times"
    - "express pass"
    - "meal plans"
    - "park maps"
    - "show times"
    - "photo pass"
    - "virtual queue"
    - "ride reservations"
    - "mobile ordering"
  
  # Pain points to prioritize
  pain_points:
    high_impact:
      - "can't book tickets"
      - "app crashes during booking"
      - "payment fails"
      - "queue times wrong"
      - "can't access tickets"
    
    medium_impact:
      - "map doesn't load"
      - "slow performance"
      - "confusing navigation"
      - "notifications not working"
  
  # Positive signals
  positive_signals:
    - "easy booking"
    - "fast check-in"
    - "helpful features"
    - "saved time"
    - "convenient"
    - "great experience"

# =============================================================================
# NOTES
# =============================================================================
# - Edit this file to customize agent behavior
# - Prompts support variables in {curly_braces}
# - Model names must match HuggingFace model IDs
# - Temperature: 0.0 = deterministic, 1.0 = creative
# - Changes take effect on next run (no restart needed for some settings)