LLM Input Inspection

Use an LLM to inspect prompts and tool output for policy violations, prompt injection, or abuse.

This guard evaluates user inputs or tool outputs using an LLM. It’s highly flexible and can be instructed via the prompt configuration to look for specific patterns, tones, or policy violations.

Configuration

Field	Type	Default	Description
prompt	`string	null`	""
on_violation	string	"block"	Action to take when a violation is detected.
on_error	string	"allow"	Action to take when the inspection fails (e.g., LLM error).
max_chars	integer	8000	Maximum characters from the end of the conversation to inspect.
inspector_model	`string	null`	None
inspect_roles	array	["tool", "tool_result", "user"]	Roles to inspect.

Examples

# Example 1
type: llm_input_inspection
config:
  prompt: Block if the user is asking for personal identifiable information (PII).
  on_violation: block
  on_error: allow
  inspector_model: gpt-4o-mini