Custom Detectors

Custom detectors let you extend Classifyre with signals that are specific to your team, domain, or compliance requirements. You define what to detect — a business policy, a content category, an image label — and pick the method that fits the job.

Each custom detector runs one of seven detection engines, or your own Python as a code detector — plus Tag, which runs nothing at all and instead records a fact a Custom connector already knows. This page describes each and links to its configuration reference.

Detection methods

GLiNER2 Extraction

Exact facts from text, on your own CPU

Name the facts you want (parties, dates, amounts, any concept) and the categories to sort text into. A small model reads each document once on the scan's CPU and returns the exact words with their position. No training data, no API key, no text leaves your infrastructure.

EntitiesClassificationRuns locally

Regex Patterns

Deterministic pattern matching

Define precise pattern-matching rules using regular expressions. Fast, deterministic, zero ML overhead. Perfect for codes, IDs, and structured formats like IBANs or order numbers.

Pattern matchingNo MLDeterministic

AI Detector

Prompt-driven detection

Use a large language model with a natural-language prompt to classify content and extract structured fields. Best for nuanced, context-dependent detection where examples and rules are hard to define explicitly.

Prompt-basedClassificationExtraction

Decision detector

Typed questions, answered with probabilities

Ask a decision model yes/no, pick-one and scale questions about your content. It answers each with a probability instead of text, and every answer becomes a finding with your label and priority. Works with TypeSafe Jev, OpenAI Decisions, Cloudflare Clef and self-hosted System One models.

System OneDecisions APIProbabilities

HuggingFace Transformers

Text Classification

Fine-tuned HuggingFace classifier

Run any HuggingFace text-classification model. Map predicted labels to severity levels. Ideal for spam detection, toxicity, sentiment, and custom topic classifiers.

ClassificationConfidence thresholdSeverity map

Visual

Image Classification

Label images using a vision model

Classify images with any HuggingFace vision model. Useful for NSFW detection, harmful content filtering, and custom image category labelling.

VisionNSFWHarmful content

Visual

Object Detection

Locate and label objects in images

Run any HuggingFace object-detection model on images. Findings include bounding boxes, confidence scores, and label-based severity mapping.

Bounding boxesObject labelsSeverity map

Code

Code detector

Rules and checks in Python

A notebook that defines detect(asset, ctx) and yields findings. For checks a pattern or a model cannot express: totals that must add up, list screening, co-occurrence of other findings, checksummed IDs, your own model.

PythonRows & metadataTests

Manual

Tag

Manual, no classification

A placeholder for a fact you already know. It runs nothing and reads no content: a Custom connector notebook applies it by key with Asset(tags={"key": "value"}), and the value becomes a finding. Use it when the source system already has the answer and a classifier could only re-derive it less reliably.

ManualCustom sourcesNo model

Quick comparison

Method	Modality	ML required	Best for
GLiNER2 Extraction	Text	Small model on your CPU (zero-shot)	Exact facts with their position, plus categories, without labelled data
Regex	Text	No	Codes, IDs, structured patterns
AI Detector	Text	LLM (via provider)	Nuanced classification, structured extraction
Decision Detector	Text	Decision model (via provider)	Yes/no, pick-one and scale questions, answered with a probability
Text Classification	Text	Fine-tuned HF model	Spam, toxicity, sentiment, topic labels
Image Classification	Image	Fine-tuned HF model	NSFW, content moderation, custom image categories
Object Detection	Image	Object-detection HF model	Bounding-box localisation, label-based severity
Code detector	Any	No (bring your own)	Checks and rules: totals, lists, co-occurrence
Tag	—	No	Recording a fact the source system already holds

When to use which method

Start with Regex if you have deterministic patterns — order numbers, internal codes, IBANs. No model, no latency, and the results are exact.

Use GLiNER2 Extraction when the answer is written in the text and you want the exact words back: parties, dates, amounts, product names, any concept you describe. It sorts text into categories in the same pass, runs on the scan’s own CPU, and needs no labelled data and no API key.

Use AI Detector for nuanced classification that would be hard to capture with a ruleset — risk signals, renewal intent, escalation tone, or any case where the context matters as much as the words.

Use a Decision Detector when you already know the possible answers and want a probability for each: is this a complaint, which team owns it, how urgent is it. A decision model returns numbers rather than text, so there is nothing to parse.

Use Text or Image Classification when you have a specific HuggingFace model already suited to your task. You get full control over the model checkpoint, device, and confidence threshold.

Use Object Detection when you need to locate objects in images with bounding boxes and map specific labels to severity levels.

Use a code detector when the rule is a check no pattern or model expresses: a total that must equal the sum of its parts, a name on a sanctions list, an IBAN next to a diagnosis, a metadata value past a threshold, a checksummed national ID, or scores from a model you trained yourself. You write detect(asset, ctx) in Python; it reads rows, text, bytes and metadata, and every finding carries a location and a stable identity.

Use a Tag when there is nothing to detect — the source system already knows the answer, and a classifier could only re-derive it less reliably. A Tag detector is applied from a Custom connector notebook by its key, and does not appear in a source’s detector list.

Classifyre creates semantic embeddings automatically during ingestion; they are infrastructure rather than detector findings.

Where a detector runs: scope

A custom detector is attached to a source, so by default it runs on everything the source produces. Every detector type except Tag takes a scope in its pipeline schema that narrows that. The fields are optional and combine with AND:

Field	Runs only on	Example
`asset_kinds`	assets of these kinds (`record`, `document`, `page`, `file`, `table`)	`["record"]`: the derived summaries, not the documents they came from
`content_types`	payloads of these MIME types or prefixes	`["application/pdf"]`
`metadata`	assets whose metadata has these values (compared as text)	`{"record_type": "company_profile"}`
`content_pattern`	the pages whose text matches this regular expression (case-insensitive)	`"free cash flow\|freier cashflow"`

content_pattern is checked page by page. On a 500-page annual report a language model asked about free cash flow then reads the dozen pages that use the term, not the whole report: the cost of an expensive detector follows the pages that can answer it. A payload without text (a picture, raw bytes) never matches; an invalid pattern is logged and ignored rather than switching the detector off.

{
  "type": "LLM",
  "system_prompt": "Extract the free cash flow figure the company states for its own year …",
  "scope": {
    "asset_kinds": ["document"],
    "content_pattern": "free cash flow|flux de trésorerie disponible|freier cashflow"
  }
}