LQL console and patches

LQL console

The console sends statements straight to the model. USE model is added automatically, so you only write the statement itself:

DESCRIBE "Acme";
INFER "The headquarters of Acme is" TOP 5;

• DESCRIBE — what the model holds about an entity.
• INFER — the model's continuation of a prompt, with TOP n alternatives ranked.

This is the direct way to check whether a fact landed, and to see what the model believed before you touched it.

Knowledge test

The Knowledge test tab wraps the same idea in a form: enter a prompt, run it, read the result. Use it to spot-check a batch after the wizard reports success.

Worth testing both directions: that the new facts are recalled, and that something unrelated the model knew before still works. That second check is what catches COMPOSE interference.

Patches

A patch is a saved set of facts for a model. The Patches tab lists them and lets you:

Action	Effect
Import patch	Load a .vlp file
Export	Write a patch out as .vlp
Delete	Remove it, after confirmation

Active patches are the ones applied when the model is queried. Ticking and unticking them is how you compare behaviour with and without a set of facts — and how you turn a bad batch off without deleting it.

Patches live in <app-data>/knowledge/patches/<model>/<name>.vlp. Because they are files, you can build a set on one machine and hand it to a colleague.

Export for LM Studio

Export for LM Studio writes the facts from the ticked patches as a Knowledge Card — a Markdown file. Paste it into the System Prompt in LM Studio, or attach it to a chat.

The card counts the facts it contains and reports where it was saved. This is the bridge out of NabuSQL: KNN recall works inside NabuSQL, and the card is how the same facts reach a model running somewhere else.

Choosing between the routes

You want	Use
Ad-hoc questions about a schema	AI assistant
A model that reliably knows a dataset, used inside NabuSQL	KNN patches
The same facts in LM Studio or another tool	Export a Knowledge Card
Facts baked into the model file itself	COMPOSE — experimental, read the warnings