Cortex AI Integration

Metatate uses Snowflake Cortex for all AI capabilities. Every AI request is processed inside Snowflake — no data leaves the consumer's account, and no external API keys or network connections are required.

Zero Egress Architecture

Traditional AI integrations require external API calls, which means data leaves your Snowflake account and traverses the public internet. Metatate takes a different approach:

Aspect	Traditional	Metatate
AI provider	External API (OpenAI, etc.)	Snowflake Cortex
Network	Outbound HTTPS to external endpoint	Internal Snowflake call
Data exposure	Prompt data sent externally	All processing inside Snowflake
Authentication	API keys stored in app	No keys needed
EAI required	Yes	No
Egress charges	Yes	No

The application uses SNOWFLAKE.CORTEX.COMPLETE() — a built-in Snowflake function that runs LLM inference within the Snowflake infrastructure. This means:

• No External Access Integrations (EAIs) to configure
• No secrets or API keys to manage
• No network egress rules or firewall exceptions
• No data processing agreements with third-party AI providers

Required Privilege

Cortex requires a single privilege on the consumer account:

IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE

This privilege is requested during application installation and grants access to the SNOWFLAKE.CORTEX functions. Without it, AI features are unavailable but all core governance functionality continues to work.

Model Selection

Metatate automatically selects the best available Cortex model in your region with automatic fallback if a model is unavailable. The selection logic is transparent to the consumer — no configuration is required.

If all models are unavailable, a structured error is returned (never a silent failure). Model availability varies by Snowflake region, and the fallback chain ensures AI features work across all supported regions.

Graceful Degradation

Metatate is designed to degrade gracefully when AI is unavailable or limited:

Missing Cortex Privilege

If the consumer has not granted IMPORTED PRIVILEGES ON DATABASE SNOWFLAKE:

• The AI status banner shows "AI Unavailable" with a clear explanation
• Natural language policy input is disabled
• AI-powered classification suggestions are disabled
• All core governance features (policy editing, compilation, deployment, materialization, MCP tools) work normally

No Available Models

If all models in the fallback chain are unavailable in the consumer's region:

• Same behavior as missing privilege
• The error message specifies that no models are available
• The system logs which models were attempted

Transient Failures

For temporary Cortex errors (timeouts, rate limits, internal errors):

• The request is automatically retried with fallback to alternate models
• If all attempts fail, a structured error is returned to the UI

What AI Powers

Feature	Requires AI	Fallback Without AI
Natural language policy input	Yes	Manual YAML editing
Classification suggestions	Yes	Manual column classification
Policy analysis and recommendations	Yes	Not available
MCP tool responses	No	Governance data served directly from tables
Policy compilation	No	Deterministic pipeline
Governance materialization	No	Deterministic pipeline
Decision engine	No	Rule-based evaluation

The core value proposition — defining, compiling, deploying, and enforcing governance policies — does not depend on AI. Cortex enhances the experience but is not required for governance operations.

How AI Is Used

Metatate sends only metadata and policy context to Cortex, never raw consumer data:

• Policy analysis: Policy YAML, table names, column names, data types
• Classification: Column names, data types, sample patterns (not row values)
• Natural language: User's natural language input describing desired policy

All prompts use structured templates that limit the context to governance metadata.