Data Security FAQ

What is the architecture of Vanna v2.0?

Vanna v2.0 is a modular agent framework built on clean abstractions:

• Core Agent: Orchestrates LLM interactions with tool execution loops, conversation management, and streaming support
• Tool System: Extensible tool registry with group-based access control
• Storage Layer: Abstract interfaces for conversations, audit logs, and observability data
• User Management: User resolution with group-based permissions (RBAC)
• LLM Services: Pluggable integrations for Anthropic Claude, OpenAI GPT, and other providers

The framework provides 6 extensibility points: lifecycle hooks, middlewares, error recovery, context enrichers, conversation filters, and observability providers.

What are the premium backend services?

Vanna's premium backend (written in Go) provides managed services:

• Observability: Metrics and distributed tracing for monitoring agent performance
• Audit Logging: Centralized audit event storage with query capabilities
• Tool Registry: Shared tool schemas and templates across teams
• Agent Memory: Semantic search over historical tool usage patterns
• Conversation Management: Cloud-based conversation persistence
• Analytics: Dashboard aggregations and usage statistics

What data does the open-source Python package handle?

When running self-hosted, the Python package handles:

1. Database Connections: Connection strings and credentials (stored locally, never transmitted)
2. Training Data: DDL statements, documentation, SQL examples, and question-SQL pairs
3. Conversation History: User messages and AI responses
4. Tool Execution Data: Tool invocations, parameters, and results
5. Audit Logs: Security events, access checks, and tool usage

What data is sent to Vanna's premium services?

When using premium backend services (opt-in), the following data may be transmitted via HTTPS:

Authentication: All premium API requests use Authorization: Bearer {api_key} headers and X-Organization-ID for multi-tenancy isolation.

How are database credentials managed?

Database credentials are:

1. Stored Locally: Credentials remain in your Python environment (environment variables, config files, or passed programmatically)
2. Never Transmitted: Credentials are never sent to Vanna servers, premium backend, or LLM providers
3. Not Logged: Audit logging automatically sanitizes sensitive parameters (password, secret, token, api_key, credential, etc.) before recording events

What user isolation and access control mechanisms exist?

Vanna v2.0 implements group-based access control (RBAC):

class User:
    id: str                          # Unique user identifier
    username: str
    email: str
    group_memberships: List[str]     # e.g., ["admin", "analyst", "viewer"]

What observability and monitoring capabilities exist?

The framework includes built-in observability:

Providers: Local (in-memory), Premium (cloud-based), or Custom (implement ObservabilityProvider for Datadog, Prometheus, etc.)

What extensibility points exist for custom security controls?

Vanna v2.0 provides multiple integration points for custom security:

What security hardening is recommended for production?

How can I ensure GDPR/CCPA compliance?

Self-Hosted Deployments: You have full control and responsibility for compliance.

• Data Minimization: Only collect necessary data
• Right to Access: Implement endpoints to export user data
• Right to Erasure: Implement deletion via conversation store APIs
• Data Retention: Configure automatic cleanup of old conversations
• Privacy Policy: Clearly disclose what data is sent to LLM providers