Snowflake Cortex

Exposes foundation models (Claude, GPT-4, Llama, Mistral) as SQL functions callable directly within Snowflake queries, eliminating data movement by executing inference inside the data warehouse boundary. Models are accessed via Snowflake's managed serverless endpoints rather than direct API calls, with results returned as SQL result sets for immediate downstream processing.

Cortex AI Functions support multimodal inputs beyond text, enabling image analysis, audio transcription, and cross-modal reasoning within SQL queries. Implementation details on how images/audio are ingested, encoded, and routed to appropriate model backends are not documented, but the capability suggests Snowflake handles format conversion and model selection internally.

Cortex integrates observability into Snowflake's monitoring and governance framework, providing visibility into LLM function execution, resource consumption, and costs. The system tracks which models are invoked, how much compute is consumed, and how results are used downstream — though specific metrics, dashboards, alerting capabilities, and cost optimization tools are not detailed.

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.

Cortex Analyst translates natural language questions into executable SQL queries, enabling non-technical users to query data without writing SQL. The system likely uses an LLM fine-tuned or prompted with schema context to generate queries, though the exact prompt engineering approach, schema inference mechanism, and query validation strategy are not documented.

Cortex Search combines text embeddings (semantic search) with traditional keyword matching to enable hybrid retrieval over unstructured data. The system automatically generates embeddings for indexed documents, stores them in a managed vector index, and routes queries to both semantic and keyword search paths, merging results via an undocumented ranking algorithm. No details on embedding model selection, index structure, or search latency are provided.

Cortex Agents coordinate multi-step workflows across structured tables and unstructured documents, routing queries to appropriate data sources and combining results. The agent likely uses an LLM to decompose user requests into sub-tasks, execute SQL queries and semantic searches, and synthesize results — but the exact orchestration logic, tool selection mechanism, and error recovery strategy are not documented.

Cortex supports fine-tuning foundation models on proprietary data and deploying custom models, though implementation details are minimal in available documentation. The capability likely involves uploading training data, configuring hyperparameters, and deploying fine-tuned models as SQL-callable functions — but the exact training infrastructure, supported model architectures, and deployment process are not specified.

Cortex executes LLM and ML workloads on Snowflake's serverless compute infrastructure, automatically scaling resources based on demand without requiring users to provision or manage clusters. Pricing is consumption-based (per-credit model), though specific per-request, per-token, or per-second costs are not disclosed. The underlying compute hardware specifications, scaling policies, and cold-start latency are not documented.

Cortex processes data within Snowflake's secure perimeter, ensuring that sensitive data never leaves the warehouse for external LLM API calls. The system enforces data governance policies (encryption, access controls, audit logging) throughout the AI pipeline, though specific encryption algorithms, audit log formats, and policy enforcement mechanisms are not documented.

Snowflake Marketplace hosts 3,400+ listings of ready-to-use datasets, applications, and services that can be directly accessed or deployed within Cortex. The marketplace enables one-click access to third-party data providers and pre-built AI applications without manual integration or data staging, though the exact listing types, deployment mechanisms, and revenue sharing models are not detailed.

Cortex provides SQL-callable access to multiple foundation models from different providers (Anthropic Claude, OpenAI GPT-4, Meta Llama, Mistral Large 2) through a unified interface, abstracting away provider-specific API differences. Users select models via SQL function parameters, and Snowflake routes requests to the appropriate provider's endpoint — but model version selection, provider failover, and cost allocation across providers are not documented.