AI & ML Availability and Settings

Gnok runs the service, its model runtimes, the AI and embedding providers, and the compute behind them. You control SQL model definitions, grants, and, as an organization administrator, your organization's AI Governance. There is nothing to install or configure on a server.

Native model training

Define algorithm options in CREATE MODEL ... OPTIONS (...). Match the declared feature and result types to the model. Training has service-managed resource limits; partitioning does not remove the need to materialize the training input. See distributed training.

ONNX Model Inference

ONNX inference is built into the hosted service. Register a compatible artifact with CREATE MODEL ... USING FRAMEWORK 'onnx'; there is no runtime library to install. Confirm the tensor contract and evaluate predictions before activation.

PyTorch and sklearn

These inference paths require the corresponding runtime to be enabled for your account. Ask Gnok support about availability before exporting an artifact for them. Model format and preprocessing requirements are described in runtime inference.

Feature Store

Feature storage is enabled by default in the engine, but effective availability and capacity are determined by the hosted service. Define the feature query, keys, output, and supported refresh settings in SQL. Background execution requires a tenant service identity with source and output grants; enabling the feature alone does not supply those permissions. See feature store.

AI Governance

AI features are off for each organization until an organization administrator enables them. This covers the AISQL functions (AI_*), RAG answers, ASK and SUGGEST QUERIES, and dashboard AI. They use Gnok's AI provider, so you don't supply a provider API key. Instead, an organization administrator opens Operations → AI Governance in Studio and does two things:

1. Saves a monthly token budget as Available. Enter a Monthly token limit, set Control state to Available, and select Save budget.
2. Enables a rollout policy. Allow the provider (anthropic), the model (claude-haiku-4-5-20251001), and the operations your organization uses, such as ai_complete or ai_classify_text. Enter the egress and retention approval references, check Egress approved and Retention approved, check Enabled, and select Save rollout policy.

Until both are in place, every AI call is refused; this is fail-closed by design. Non-AI queries are unaffected. Each organization sets up its own budget and policy. The tutorial setup lists the operations each AI tutorial needs.

Natural Language Interface

ASK and SUGGEST QUERIES run under the same AI Governance: the budget must be Available, and the rollout policy must allow the provider and model, with egress and retention approved. EXPLAIN ASK makes no provider call. Select a focused catalog/schema context and inspect generated SQL before execution.

AI scalar functions

Each AI_* call needs an Available token budget and an enabled rollout policy that allows the function's operation (for example, ai_summarize for AI_SUMMARIZE). Concurrency, deadlines, and retry limits are managed by Gnok. Failed attempts that reach the provider may still use tokens. See AI functions for nullable variants and error behavior.

Embedding

EMBED() uses Gnok's embedding provider within your organization's monthly embedding allowance. It doesn't need AI Governance. Document and query embeddings must use the same model and vector dimension. The current SQL EMBED path declares a 1536-dimensional Float32 vector; a different provider model name alone does not change that output type. See vector operations.

Vector Indexes

Use index DDL for construction options and documented index or session settings for search parameters. Confirm that the index is visible in the intended catalog/schema and inspect results for the qualified name. See vector operations.

Learned Optimization

Optimizer feature enablement and training policies are service-managed. Use the documented diagnostics to compare estimates and actual results. Learned corrections do not guarantee a faster plan for every query. See learned optimization.

AutoML Advisors

Advisor availability and automatic application are separate service capabilities. Review recommendations and permitted actions with your administrator; an advisor being available does not mean it may automatically modify your data or resources. See AutoML advisors.

Request access or diagnose a failure

If AI calls are refused, an organization administrator checks AI Governance first. Use Studio administration for other tenant controls. Ask Gnok support about runtimes, integrations, or capacity that your account doesn't have. Include the organization name and the error reference, never credentials. Tutorial setup lists workload-specific requirements.