Cortafuegos de IA
El cortafuegos de IA es una capa de seguridad diseñada específicamente para proteger los sistemas de inteligencia artificial, especialmente los modelos de lengu...
9 min de lectura
AI Security
LLM
+3
The Custom Guardrail component functions as an intelligent filter within an AI workflow, ensuring that user input aligns with specific criteria set by the workflow designer. This component leverages a language model (LLM) in combination with a customizable prompt to validate and control the flow of conversation or data.
What the Component Does
At its core, the Custom Guardrail serves as a gatekeeper for user inputs. It uses a prompt (the “Guardrail Prompt”) to instruct the LLM on what is considered acceptable content. For example, the default prompt restricts conversations strictly to sports topics and blocks any unrelated or nonsensical (gibberish) inputs. This mechanism ensures that the AI workflow remains focused and does not deviate from its intended purpose.
How It Works
- The component receives an input text (typically a user message).
- It sends this input, along with the guardrail prompt, to the selected LLM.
- The LLM evaluates the input based on the prompt and determines whether it meets the criteria.
- Depending on the result, the component outputs the message along one of two routes: accepted (true) or rejected (false).
Inputs
| Name | Type | Description |
|---|---|---|
| LLM | HandleInput | The language model used to evaluate and enforce the guardrail. |
| Guardrail Prompt | PromptInput | The instruction provided to the LLM on what content is allowed. |
| Input Text | Message | The user’s message or content to be validated. |
- LLM: Choose the model you wish to use as your guardrail enforcer. This allows flexibility in using different LLMs suited for your needs.
- Guardrail Prompt: Customize how strict or lenient the guardrail should be by editing the prompt.
- Input Text: The actual text that needs validation.
Outputs
| Name | Type | Description |
|---|---|---|
| True Route | Message | Output if the input text passes the guardrail check (i.e., meets the criteria). |
| False Route | Message | Output if the input text does not meet the criteria and is filtered out/rejected. |
- True Route: When the input complies with the guardrail, it is passed through this output for further processing.
- False Route: If the input is off-topic or fails the check, it is sent through this output, allowing you to handle it differently (e.g., send a warning or request clarification).
Why Use This Component?
- Quality Control: Enforce strict topic boundaries or content rules within your AI applications.
- Safety and Relevance: Prevent AI from responding to off-topic, irrelevant, or inappropriate content.
- Customizable Enforcement: Easily modify the guardrail prompt to reflect changing requirements or policies.
- Seamless Integration: Works with a variety of LLMs and accepts flexible input/output routing in your workflow.
Use Cases
- Moderating chatbots to stay on-topic (e.g., customer support, educational bots).
- Filtering user-generated content before it is processed or stored.
- Enforcing compliance or legal requirements in automated AI interactions.
- Preventing spam or nonsensical inputs from disrupting workflows.
By integrating the Custom Guardrail component into your AI workflow, you gain fine-grained control over what content is accepted, helping ensure your AI system behaves as intended and delivers reliable, relevant results.