Websocket message validation

The LLM Websocket message validator plugin uses an LLM to validate WebSocket messages in real-time. Each text message sent by the client is checked against an LLM prompt before being forwarded to the backend.

How it works

1. A WebSocket connection is established between the client and the backend through Otoroshi
2. When the client sends a text message, the plugin intercepts it
3. The message content is sent to the configured LLM provider with a system prompt
4. The LLM returns "true" (valid) or "false" (invalid), optionally with a reason
5. Valid messages are forwarded to the backend. Invalid messages are dropped or replaced with an error message
6. Binary messages pass through without validation

Plugin configuration

{
  "enabled": true,
  "plugin": "cp:otoroshi_plugins.com.cloud.apim.otoroshi.extensions.aigateway.plugins.AiWebsocketMessageValidator",
  "config": {
    "ref": "provider-entity-id",
    "prompt": "You are a content moderator. Analyze the following WebSocket message. Return 'true' if the message is appropriate, or a JSON object {\"result\": false, \"reason\": \"explanation\"} if it should be blocked.",
    "prompt_ref": null,
    "context_ref": null,
    "extractor": null
  }
}

Parameters

Parameter	Type	Default	Description
ref	string	""	LLM Provider entity ID
prompt	string	""	System prompt instructing the LLM how to validate messages
prompt_ref	string	null	Reference to a stored prompt entity
context_ref	string	null	Reference to a stored context entity
extractor	string	null	Regex pattern to extract the result from the LLM response

LLM response format

The LLM can return:

• "true" — message is valid, forward to backend
• "false" — message is invalid, drop it
• {"result": false, "reason": "explanation"} — message is invalid, send the reason back to the client as a text message

Example: chat moderation

{
  "ref": "provider_openai",
  "prompt": "You are a chat moderator. Analyze the following message and determine if it is appropriate for a professional chat room. Return {\"result\": true} if appropriate, or {\"result\": false, \"reason\": \"brief explanation\"} if not. Block profanity, harassment, spam, and off-topic content."
}

Example: data validation

{
  "ref": "provider_openai",
  "prompt": "The following is a WebSocket message that should contain valid JSON with a 'type' and 'data' field. Return 'true' if it is valid, 'false' if not."
}