Skip to main content
This tutorial covers @xpert-ai/plugin-sensitive-filter from xpert-plugins, used to filter sensitive content on both agent input and output stages.

​
Typical Use Cases

  • Filter PII such as ID numbers, phone numbers, and bank cards
  • Block high-risk content directly
  • Rewrite recoverable content into a safe response

​
Install and Enable

  1. Install the plugin package in your host project:
npm install @xpert-ai/plugin-sensitive-filter
  1. Enable the plugin through environment variables:
PLUGINS=@xpert-ai/plugin-sensitive-filter
  1. Follow Publish & Use to ensure the host loads the plugin.

​
Runtime Hooks

  • beforeAgent: evaluate and optionally rewrite/block input
  • wrapModelCall: evaluate and optionally rewrite/block model output
  • afterAgent: write audit snapshot

​
Configuration Modes

The middleware has two mutually exclusive modes:
  • rule: deterministic rules (keyword / regex)
  • llm: natural-language policy evaluation (LLM hits are enforced in rewrite behavior)

​
Minimal Rule Mode Example

{
  "mode": "rule",
  "caseSensitive": false,
  "normalize": true,
  "rules": [
    {
      "id": "rule-1",
      "pattern": "ID card",
      "type": "keyword",
      "scope": "both",
      "severity": "high",
      "action": "block",
      "replacementText": "Sensitive content was blocked."
    }
  ]
}
In rule mode, each rule should include: pattern, type, scope, severity, and action.

​
Minimal LLM Mode Example

{
  "mode": "llm",
  "llm": {
    "model": {
      "provider": "openai",
      "model": "gpt-4o-mini"
    },
    "scope": "both",
    "rulePrompt": "If content contains ID cards, phone numbers, bank cards, or home addresses, rewrite it into a privacy-safe response.",
    "rewriteFallbackText": "[Filtered]",
    "timeoutMs": 3000
  }
}
In llm mode, runtime-required fields are: model, scope, and rulePrompt.

​
Validation Checklist

  1. Validate the hit path in rule mode first.
  2. Then switch to llm mode and validate semantic policy behavior.
  3. Verify that audit records include hit details for both input and output phases when expected.

​
Troubleshooting

  • No effect in rule mode: usually caused by incomplete rule fields or mismatched scope.
  • No effect in llm mode: confirm model, scope, and rulePrompt are all present.
  • Unexpected LLM rewrite behavior: inspect audit traces for policy fallback/error hints.