Filter recipes

The gear icon (Administration) › Filter recipes is where you, as an admin, choose which recipes new filters may use on this deployment. Every recipe is on until an admin turns it off. Turning one off changes what can be built next; it never stops a filter that is already running.

What the page shows

The page has one panel per pipeline slot, in the order a pipeline runs them: Always escalate, Collapse repeats, Decide and Shape delivery. Each recipe is listed under every slot it can fill. A recipe in two slots (Fixed rule) has one switch, shown in both rows; turning either turns both.

Slot	Recipes	Recipe’s own minimum labels
Always escalate	Fixed rule (`rule_prefilter`), Plain-language rules (`decision_rules`)	0
Collapse repeats	Collapse repeats (`dedup_cluster`)	0
Decide	Fixed rule (`rule_prefilter`)	0
	Text analysis (small sample) (`text_clf_setfit`)	50
	Text analysis (`text_clf_embed_lr`)	300
	Anomaly detection (`anomaly_novelty`)	0
	Structured field analysis (`tabular_gbdt`)	500
Shape delivery	Group escalations (`digest`)	0

A filter’s first training waits for the higher of its Min labels to train setting (400 by default) and the recipe’s minimum, so even a recipe that needs 0 waits for that; see the label floor.

Each row shows the recipe’s title, a chip for the recipe’s own minimum (and for an optional extra it needs installed, such as setfit), and one of three states:

Badge	Meaning	Control
ready	This install can run it and it is on.	Switch, On
off	An admin turned it off for new filters.	Switch, Off
missing its requirements	This install cannot train it. The row says why and What to do.	None: “nothing to decide here”

The ? beside a title gives the recipe’s summary, what it needs, and its recipe id.

Why a recipe can be missing its requirements

Two recipes depend on the install rather than on your choice.

Text analysis (small sample) needs an optional larger install, which an operator sets up. Without it the row says the image “was built without the optional ‘setfit’ extra”. Ask your operator; the steps are in Install without internet access and Configure Sift. Text analysis needs neither.
Plain-language rules needs a decision model configured. Without one the reason begins “Needs a decision model.”

These are checked each time the page loads, so after you change the image or stage the weights, reload the page; nothing else needs restarting.

Turn a recipe off

Select the recipe’s switch (On).
In Turn off {recipe}?, read the consequence and select Turn it off.

The badge reads off. Turning a recipe back on is one click and asks nothing, since it restores the default.

What turning one off does

Affected	Not affected
Sift’s assistant no longer proposes the recipe for a new filter.	Filters already deciding with it keep scoring.
It cannot be added as a candidate in the slate editor.	They keep retraining.
The server refuses a filter version that would newly train it.	They can still save versions that change something else.

So turning a recipe off is a statement about new work, not a kill switch. To stop a filter that uses it, hold, stop or retire the filter.

How an operator sees it

In the new-filter flow, a proposal that uses a recipe turned off here shows a warning before anything is saved: “{recipe} is turned off for new filters. Choose another in {slot}, or change what the reason names.”, followed by the reason “An administrator turned this recipe off for new filters.” and What to do: “Turn it back on under Administration > Filter recipes.” A candidate in the slate editor reads “It is turned off for new filters; uncheck it to go on.”

The same messages name the other causes in their own words: a recipe that “cannot be trained on this install” (with the image or weights to fix), or one that “needs a decision model”.

You cannot turn on a recipe the install is missing the requirements for; the server refuses it, and the screen offers no switch.

Audit

Every change writes recipe.enabled or recipe.disabled to the audit log, with the recipe and its state before and after. Who last changed a recipe is read there, not on this page. A request that changes nothing writes nothing.