Detect PII

Vela 1.0 PII finds personal information in a request and returns where it is: the type, the exact characters and a probability for every span. A pii signal matches when it finds a type you do not allow, and a route can then keep the request on a private model or refuse it.

The model knows 17 types: AGE, CREDIT_CARD, DATE_TIME, DOMAIN_NAME, EMAIL_ADDRESS, GPE (places), IBAN_CODE, IP_ADDRESS, NRP (nationality, religion or political group), ORGANIZATION, PERSON, PHONE_NUMBER, STREET_ADDRESS, TITLE, US_DRIVER_LICENSE, US_SSN and ZIP_CODE.

Turn it on

routing:
  signals:
    pii:
      - name: personal_data
        threshold: 0.85
        pii_types_allowed: [EMAIL_ADDRESS]
  decisions:
    - name: keep-private
      priority: 200
      rules:
        operator: AND
        conditions:
          - type: pii
            name: personal_data
      modelRefs:
        - model: private-model

Without an override, the Router uses the default Vela 2.0 deployment for PII. Its input limits and coverage differ from the Vela 1.0 specialist below; see Choose a model. Rule thresholds are explicit policy: calibrate this example's value for the model you select.

The Vela 1.0 binding below scans up to 32,768 tokens in overlapping 512-token windows. Coverage includes the end of the document, but detection accuracy still depends on the text and model.

Choose where it runs

The binding is pii_classifier and reads text spans (token_spans.v1):

global:
  model_catalog:
    deployments:
      vela-pii:
        provider: model_runtime
        artifact: vllm-sr/Vela-1.0-Encoder-307M-PII
        device: cpu
        input:
          max_tokens: 32768
          overflow: window
    bindings:
      pii_classifier:
        deployment: vela-pii
        contract: token_spans.v1
    modules:
      classifier:
        pii:
          window:
            size: 512
            overlap: 255

Keep overflow: window for PII: the model reads the whole request in 512-token windows that overlap by 255 tokens. truncate would scan only the beginning, and reject makes the signal unknown for longer requests.

On Vela 2.0

Vela 2.0 finds the same 17 types with its router span head, as a ready-made question. Bind pii_classifier to a Vela 2.0 deployment and the PII question travels in the same call as the decision questions the request asks that deployment about the same text:

global:
  model_catalog:
    deployments:
      vela2:
        provider: model_runtime
        artifact: vllm-sr/Vela-2.0-0.3B
        device: cpu
    bindings:
      pii_classifier:
        deployment: vela2
        contract: token_spans.v1

Vela 2.0 uses its own full-input checks and supported scan budget; it does not use the specialist's 512-token PII window settings. An input override can set the deployment's scan budget. Incomplete coverage remains unresolved; see Long inputs and the pii signal for how rule thresholds apply to its spans.

When the scan cannot finish

If the model is not ready or a scan fails, the signal is unknown. Content the model did not read (over its input or scan cap, truncated, or not scanned by the signals' deadline) matches as unscanned unless the module sets on_unscanned: allow. The PII module's on_error decides what an unknown signal means: allow (default) treats the unread text as clean, and block matches it as classification_error, so text that could not be checked cannot pass as clean.

global:
  model_catalog:
    modules:
      classifier:
        pii:
          on_error: block

Check it

These worker-level examples run inside an environment containing vllm-srun (such as the Router image). Classify, embeddings, rerank and bundle are worker APIs; the instance frontend publishes System One and decision requests.

vllm-srun serve vllm-sr/Vela-1.0-Encoder-307M-PII --device cpu --port 8100
curl -s localhost:8100/v1/classify -H 'content-type: application/json' \
  -d '{"input": ["Hi, I am Tom Baker, write to tom.baker@example.com."]}'

Each span has its label, start and end (in characters, end exclusive), the text and its probability. Through the router, x-vsr-matched-pii lists the PII rules that matched.

On Vela 2.0, ask the ready-made question on /v1/decisions; the spans are in spans.pii:

curl -s localhost:8100/v1/decisions -H 'content-type: application/json' -d '{
  "state": "Hi, I am Tom Baker, write to tom.baker@example.com.",
  "questions": {"pii": {"preset": "pii"}}
}'