Skip to main content

Prompt Classifier – Admin Guide

Purpose​

Classify prompts, export an allow‑listed category to PDP for policy/budgeting, and optionally enable a strict guard.

Configuration​

  • File: ServiceConfigs/BFF/config/classifier.yaml
  • Optional overrides: ServiceConfigs/BFF/config/classifier.d/*.yaml

Key fields

  • enabled: true|false
  • backend: heuristic|hf|hf_zero_shot|onnx
  • model: <hf model id> (for ML backends)
  • labels: [ ... ]
  • policy_export:
    • enabled: true
    • min_conf: 0.60
    • labels_allowlist: [ "salary", "payroll", ... ]
  • guard (strict, optional):
    • enabled: true
    • block_labels: [ ... ]
    • min_conf: 0.50

Runtime controls​

  • Header: X-Category-Mode: strict|lazy (strict → guard may block; lazy → advisory/export only)
  • Env override: CATEGORY_EXPORT_ENABLED=true|false

Rollout tips​

  • Start with export only (lazy); keep guard off in prod initially.
  • Canary allowlist; align with PDP confidence/labels.

See also​

  • Budgets (selectors and 402 mapping): services/bff/how-to/llm-routing-budgets.md
  • PDP Effective Budgets: services/pdp/reference/effective-budgets.md