Tag

Manual · Custom sources · No model

A Tag detector runs nothing. It loads no model, reads no content, and can never find anything on its own. It exists to give a fact you already know a stable identity — a key, a label, a severity — so that a Custom connector can record it and have it behave like any other finding.

A Tag detector is only usable from a Custom (notebook) connector. It does not appear in a source’s detector list, and nothing else can trigger it.

When to use

Use a tag when the source system already holds the answer and running a classifier over the content could only re-derive it less reliably.

A catalog classification — the data catalog already says this table holds cardholder data. Copy the fact; do not try to re-detect it.
A legal hold or retention flag — something a person decided, recorded upstream, and that appears nowhere in the content.
Jurisdiction or ownership — where a dataset came from, or which team owns it, when that lives in metadata rather than text.
An upstream review outcome — a prior audit’s verdict you want visible next to what Classifyre finds itself.

Do not use a tag when the answer is in the content and has to be found. That is what the other engines are for.

Reach for	When
Tag	The source system already holds the answer.
Regex	The answer is a fixed, structured token in the content.
GLiNER2	An entity or category has to be extracted, with no labelled data.
AI Detector	The answer needs judgement about context or intent.

How it works

Three steps, and no scan configuration in between.

1. Create the detector. Detectors → New → Tag. Give it a name, a key, a label and a severity. The key is the only part your notebook has to know.

2. Apply it in a notebook. Each key in Asset.tags is a Tag detector’s key; the value is what you are asserting.

from classifyre import Asset
 
 
def extract():
    for table in catalog.tables():
        tags = {}
        if table.classification:
            tags["cardholder_data"] = table.classification
 
        yield Asset(
            id=table.id,
            name=table.name,
            content=table.sample,
            kind="table",
            tags=tags,
        )

3. Scan. Every tag becomes a finding on that asset:

Finding field	Comes from
Detector	The Tag detector the key names.
Type	`tag:<label>`
Severity	The detector’s severity.
Matched content	The value your notebook passed.
Confidence	Always `1.0` — a tag is asserted, not inferred.

There is no step where you add the detector to a source. Every active Tag detector is available to every Custom connector automatically, so a new tag needs no source change.

A tag whose key matches no Tag detector is reported as a scan warning and skipped — the warning names the key and lists the keys that do exist. If a tag does not appear on an asset, look there first.

Configuration

Parameter	Type	Required	Description	Default	Constraints
scope	object	No	Restrict this detector to part of a source (asset kind, content type, or an asset-metadata predicate). Null runs it on everything the source produces.	null	—
label	string	No	Label carried by every finding this detector produces (finding_type is 'tag:<label>'). Defaults to the detector's name when omitted.	—	—
severity	string	No	Severity assigned to every finding this detector produces. There is no confidence to derive it from: a tag is asserted, not inferred. Allowed: critical, high, medium, low, info	medium	—

The detector’s key, name and description are set alongside the schema, as they are for every custom detector. The key is what a notebook writes; the label is what every finding is called.

Examples

Cardholder data, from a catalog classification

{
  "type": "TAG",
  "label": "Cardholder data",
  "severity": "high"
}

yield Asset(
    id="prod.payments.transactions",
    name="transactions",
    kind="table",
    tags={"cardholder_data": "primary-account-numbers"},
)

Legal hold, where severity is informational

{
  "type": "TAG",
  "label": "Legal hold",
  "severity": "info"
}

yield Asset(
    id="matter-4471/exports",
    name="Matter 4471 exports",
    kind="file",
    tags={"legal_hold": "retained-pending-litigation"},
)

Several tags on one asset

yield Asset(
    id="prod.hr.employees",
    name="employees",
    kind="table",
    tags={
        "cardholder_data": "none",
        "jurisdiction": "EU",
        "legal_hold": "retained-pending-litigation",
    },
)

Lifecycle

A tag finding lives the same life as any other finding, with one difference worth knowing.

Changing a tag’s value resolves the old finding and opens a new one. The value is part of the finding’s identity, so "EU" becoming "UK" is a different assertion, not an edited one.
Changing only a tag re-scans the asset. Tags are part of the asset’s checksum, so a value that changes while content does not still takes effect.
Deactivating a Tag detector stops it being applied. A notebook still writing its key gets the unknown-key scan warning.
Deleting a Tag detector resolves its findings, like deleting any other detector.

Not trainable

There is nothing to train. A Tag detector has no model, no examples and no inference step, so the detector page shows no training history and no Train now button. Its accuracy is exactly the accuracy of the system you copied the fact from.

Full notebook syntax is in the Notebook Reference.