Personalization Engine

The Personalization Engine learns user preferences over time and adapts AI responses to match individual interaction patterns, preferred visualization types, query complexity levels, and domain focus areas. It operates as an internal integration within the AI Service, feeding into the agent orchestrator and response generation pipeline.

Architecture

The personalization system is implemented in src/personalization/ with four components:

User Profile Model

Each user builds a profile over time through interactions:

class UserProfile:
    user_id: str
    tenant_id: str
    preferences: UserPreferences
    interaction_history: InteractionSummary
    skill_level: SkillLevel  # beginner, intermediate, advanced
    domain_expertise: dict[str, float]  # domain -> confidence
    created_at: datetime
    last_active: datetime

Preference Categories

Personalization Signals

The engine collects signals from multiple sources:

Explicit Signals

• User sets preferred chart types in settings
• User selects a persona during onboarding
• User bookmarks specific dashboards or queries

Implicit Signals

• Query topics and tables accessed
• Visualization types accepted vs. modified
• Response ratings and corrections
• Session duration and engagement patterns
• SQL complexity of accepted queries

Response Adaptation

The personalization engine influences responses at three points:

1. Agent Routing

class PersonalizedRouter:
    async def select_agent(self, query: str, profile: UserProfile) -> str:
        if profile.skill_level == SkillLevel.BEGINNER:
            return "guided_sql_agent"  # More explanatory
        return "standard_sql_agent"  # Concise responses

2. Context Assembly

The schema context retrieval is biased toward tables and columns the user frequently queries:

async def get_personalized_context(query: str, profile: UserProfile):
    base_context = await get_schema_context(query)
    # Boost relevance of frequently-used tables
    for table in profile.preferences.frequent_tables:
        base_context.boost(table, weight=1.5)
    return base_context

3. Response Formatting

Response detail level is adjusted based on user expertise:

Configuration

Privacy

Personalization profiles respect user privacy settings:

• Users can opt out of personalization entirely
• Profiles are tenant-isolated and never shared across tenants
• PII is stripped from interaction history before profile computation
• Users can request profile deletion through the Privacy Center