Skip to main content

Model-Driven Skills

Skills can reach the model three ways — always, similarity, or model-driven. The model-driven trigger is gated behind a tenant-level switch you control here, because it changes what tools QRY exposes to the model.

What the toggle does​

When Model-Driven Skills is off (the default):

  • Skills with the always and similarity triggers work normally.
  • Skills set to model-driven are not offered to the model. The catalog, load_skill, and run_script tools are not exposed.

When it's on:

  • QRY shows the model a compact catalog — each model-driven Skill's name and one-line description.
  • The model can call load_skill to pull in a Skill's full content, or run_script to execute a script Skill, only when it judges the Skill relevant to the current question.
  • Only the Skills actually used consume context-window tokens — so a large library of Skills and scripts doesn't bloat every prompt.

This is progressive disclosure: keep the base prompt small, make a big library available on demand.

Enabling it​

  1. Go to Admin → System Settings → Model-Driven Skills.
  2. Flip the switch on.
  3. Save.

The change propagates without a restart. New conversations pick it up immediately; existing conversations pick it up on their next turn.

Unified Skills must be active

The panel shows a warning if the toggle is on but the unified Skills pipeline isn't active for the tenant yet. In that state the model-driven tools won't actually be exposed even though the switch is on. The status indicator reflects the effective state (enabled and unified active), so trust the dot, not just the switch.

Security​

  • Reading: a model-driven Skill only appears in the catalog for users who can see its scope (and, for datasource-scoped Skills, who have access to that datasource).
  • Running scripts: script Skills execute in the sandboxed Python environment under the triggering user's RBAC/ABAC. A Skill can never grant access the user doesn't already have.
  • Managing: who can author or edit a Skill is governed by its scope (system/datasource admins, workspace owners, or the user themselves).

Seed Skills​

New tenants can be provisioned with built-in showcase Skills — for example a financial-analysis procedure Skill — so users have a working, editable example of the model-driven pattern from day one. These behave like any other Skill: scoped, versioned, and editable in the Skills IDE.

Troubleshooting​

A model-driven Skill never fires. Check, in order: (1) the toggle is on and the status dot shows active; (2) the requesting user can see the Skill's scope; (3) the Skill's one-line description clearly signals when it's relevant — the model decides from the description alone.

Scripts fail to run. Script Skills use the same sandbox as Python execution. Confirm the user has access to the datasource the script queries, and check the sandbox resource limits.

See also​

QRYA product of IXEN.