AI Detector

Text · LLM · Prompt-based

The AI Detector sends asset content to a configured LLM provider along with a system prompt you write. The model classifies the content against a label taxonomy you define, optionally extracts structured fields, and returns findings with the predicted labels and extracted data attached.

When to use

Nuanced classification — renewal risk, escalation intent, executive tone, regulatory concern — signals where rules and keyword lists fall short.
Structured extraction from free text — pull contract party names, renewal dates, risk clauses, or routing metadata from unstructured documents.
Rapidly evolving categories — add or rename labels by editing the config; no retraining needed.
Multi-label content — the multi_label option allows a single asset to receive multiple findings when content spans several categories.
Explanation-oriented workflows — the prompt is a plain-English artefact that stakeholders can read and approve.

How it works

Classifyre injects the asset content and your system_prompt into the LLM provider at dispatch time. The model returns a JSON response with predicted labels and any extracted fields. Each predicted label becomes a finding; extracted fields are stored in the finding’s extracted_data map.

The labels array defines the classification taxonomy. Each label has a name and an optional description that guides the model. The severity_map maps specific label predictions to severity levels. The output_fields array defines structured properties the model should extract alongside classification.

The LLM provider and credentials are configured at the source level — the detector config itself is provider-agnostic. See AI Provider configuration for setup details.

Configuration

Parameter	Type	Required	Description	Default	Constraints
system_prompt	string	Yes	Instruction describing what the model should detect, classify, and extract.	—	—
response_example	string	No	Optional few-shot example of the JSON the model should return.	null	—
temperature	number	No	Sampling temperature. Lower is more deterministic.	0	min 0, max 2
max_tokens	integer	No	Maximum tokens to generate. Provider default when null.	null	—
labels	array	No	Classification taxonomy the model assigns to content.	[]	—
multi_label	boolean	No	Allow more than one label per asset.	false	—
severity	string	No	Default severity when no severity_map rule matches a predicted label.	info	—
severity_map	array	No	Ordered rules mapping predicted labels to severity levels. First matching rule wins.	null	—
confidence_threshold	number	No	Minimum model confidence to report a label as a finding (0-1).	0.5	min 0, max 1
output_fields	array	No	Structured properties the model extracts. Stored in finding metadata and extracted_data.	[]	—
content_limit	integer	No	Maximum characters of content sent to the model.	8000	min 1
provider_runtime	object	No	Runtime-only credentials injected by the API at dispatch. Never persisted; rejected on create/update.	null	—

Examples

Renewal risk classifier with field extraction

{
  "type": "LLM",
  "system_prompt": "You are a contract analyst. Classify the document for renewal risk and extract the key renewal date and responsible party.",
  "labels": [
    { "name": "no_risk", "description": "No renewal concerns identified" },
    { "name": "low_risk", "description": "Mild risk signals, monitoring recommended" },
    { "name": "high_risk", "description": "Clear renewal risk, action required" }
  ],
  "severity_map": [
    { "label": "high_risk", "severity": "high" },
    { "label": "low_risk", "severity": "medium" }
  ],
  "output_fields": [
    { "name": "renewal_date", "description": "The contract renewal or expiry date" },
    { "name": "responsible_party", "description": "Person or team responsible for renewal" }
  ],
  "severity": "info",
  "confidence_threshold": 0.6
}

Executive escalation detector

{
  "type": "LLM",
  "system_prompt": "You are a support triage assistant. Detect whether this message requires executive escalation based on language, urgency, or explicit requests.",
  "labels": [
    { "name": "no_escalation", "description": "Standard support request" },
    { "name": "escalate", "description": "Requires manager or executive attention" }
  ],
  "severity_map": [{ "label": "escalate", "severity": "high" }],
  "temperature": 0.0,
  "max_tokens": 256
}