Custom detectors

Custom detectors let you catch data that only your organization cares about: customer account numbers, internal project code names, part numbers, ticket IDs, unreleased product names. Define them in YAML, test them with the CLI, and refer to them in policies like built-in data classes.

Regex with a validator

apiVersion: promptglass.ai/v1
kind: Detector
metadata:
  name: acme-customer-account
spec:
  dataClass: acme.customer_account
  description: "Acme customer account numbers (AC-########-C)"
  pattern: '\bAC-\d{8}-[0-9X]\b'
  validator:
    type: mod11              # luhn | mod11 | mod97 | iban | custom-wasm
  context:
    words: [account, customer, acct, "client id"]
    window: 60               # characters on either side
    required: false          # when true, a match without context is dropped
    boost: high              # confidence when context is present
  allowlist: ["AC-00000000-0"]

Dictionary

kind: Detector
metadata: { name: project-codenames }
spec:
  dataClass: acme.codename
  dictionary:
    caseSensitive: false
    wholeWord: true
    terms: [Bluefin, Kestrel, "Project Atlas", Nightjar]
  minCount: 1

Dictionaries can hold hundreds of thousands of terms. They compile to an Aho-Corasick automaton, so lookup cost doesn’t grow with dictionary size.

Keyword proximity

Finds documents that talk about a sensitive topic without fixed identifiers:

kind: Detector
metadata: { name: unreleased-earnings }
spec:
  dataClass: business.financials.unreleased
  proximity:
    all: [[revenue, EBITDA, "gross margin", guidance], [Q1, Q2, Q3, Q4, FY2026, FY2027]]
    any: [draft, "not for distribution", preliminary]
    window: 200

Custom ML classifier

Upload labeled examples (at least 50 positive and 50 negative) under Detection → Classifiers → New. PromptGlass fine-tunes a small text classifier inside your tenant. Your examples never leave your deployment, and they are never used to train anyone else’s models. Refer to the classifier as classifier: <name> in a detector.

Test and apply

promptglass detector apply -f detectors/
promptglass detector test acme-customer-account --file samples/ticket.txt
promptglass detector test --all --corpus ./regression-corpus/   # precision/recall report

Keep detector files in Git next to your policies. A CI job that runs promptglass detector test --corpus before apply catches regressions early.