Snowflake Cortex

Exposes foundation models (Claude, GPT, Llama, Mistral) as SQL functions callable directly within Snowflake queries without leaving the data cloud. Requests are routed through Snowflake's managed serverless compute layer, which handles authentication, rate limiting, and response streaming back into result sets. This eliminates the need for external API calls, data export, or custom orchestration code.

Provides built-in vector indexing and approximate nearest neighbor (ANN) search within Snowflake tables, enabling semantic search over embeddings without external vector databases. Vectors are stored as native Snowflake VECTOR data types, indexed automatically, and queried via SQL functions. Supports similarity metrics (cosine, Euclidean) and integrates with Cortex's embedding models to generate vectors from text or images in-place.

Enables deployment of Cortex operations across multiple Snowflake regions while maintaining data residency compliance. All LLM calls, embeddings, fine-tuning, and vector search operations execute within the specified region, ensuring data never crosses regional boundaries. Supports failover and disaster recovery in Business Critical edition, with automatic replication of models and indexes across availability zones.

Enables deployment of trained ML models (including fine-tuned LLMs) as SQL functions, making inference callable directly from SQL queries without external APIs or application code. Supports batch inference on large datasets, real-time inference in stored procedures, and integration with Snowflake's query optimizer for efficient execution. Models are versioned and can be rolled back or A/B tested within SQL.

Generates dense vector embeddings from text, images, and audio files using Cortex-hosted embedding models, storing results as VECTOR data types in Snowflake tables. Embeddings are computed serverlessly within Snowflake's infrastructure and can be immediately indexed for semantic search or used as features for downstream ML models. Supports batch processing of large datasets without data export.

Enables fine-tuning of supported foundation models (exact list not documented) using custom datasets stored in Snowflake tables. Fine-tuning jobs are executed serverlessly within Cortex's managed infrastructure, and resulting models are deployed as SQL-callable functions. Supports supervised fine-tuning for classification, summarization, and generation tasks without requiring external ML platforms.

Provides a framework for building autonomous agents that decompose complex tasks into multi-step workflows, coordinate between LLMs and SQL queries, and maintain state across interactions. Agents can plan, execute SQL queries, retrieve context from vector search, and iterate based on results—all within Snowflake's governance boundary. Supports agent-to-agent communication and integration with external tools via function calling.

Enables analysis of unstructured data (documents, PDFs, images, transcripts) stored in Snowflake STAGE or as binary columns using Cortex's LLM and vision capabilities. Supports document parsing, OCR, entity extraction, and content summarization via SQL functions. Processed results are stored back in Snowflake tables for downstream analysis, search, or reporting without data export.

Enforces role-based access control (RBAC) and data governance policies on all Cortex operations, ensuring that LLM function calls, vector searches, and model deployments respect Snowflake's existing security model. Integrates with Snowflake's Dynamic Data Masking (DDM) and Row Access Policies (RAP) to prevent unauthorized data exposure. Audit logs track all AI operations for compliance and cost attribution.

Provides serverless, auto-scaling compute infrastructure for all Cortex operations (LLM calls, embeddings, fine-tuning, vector search) without requiring users to provision or manage hardware. Compute is allocated on-demand based on workload, scaled automatically to handle spikes, and billed based on consumption (exact pricing model not documented). Supports both on-demand and pre-paid capacity options for cost optimization.

Integrates Cortex with Snowflake Marketplace, enabling users to access 3,400+ pre-built datasets, applications, and (potentially) AI models without data movement. Marketplace listings can be directly queried in SQL or used as context for Cortex LLM functions. Supports live data access from third-party providers (750+ business and data providers) for real-time enrichment and analysis.

Provides consumption tracking and cost attribution for all Cortex operations (LLM calls, embeddings, fine-tuning, vector search) using Snowflake's credit-based billing model. Tracks usage by operation type, user, role, and warehouse, enabling cost allocation and budget management. Offers cost optimization tools and visibility into consumption patterns without requiring external cost monitoring solutions.