Topical Guard

The Topical Guard is an input guard that ensures user inputs adhere to a predefined set of allowed topics. It analyzes the input, compares it against the specified topics, and determines whether the input is relevant or off-topic.

info

TopicalGuard is only available as an input guard.

Example

from deepeval.guardrails import TopicalGuard

allowed_topics = ["technology", "science", "health"]
user_input = "Can you tell me about the latest advancements in quantum computing?"

topical_guard = TopicalGuard(allowed_topics=allowed_topics)
guard_result = topical_guard.guard(input=user_input)

There is 1 required parameter when creating a TopicalGuard:

• allowed_topics: A list of topics (strings) that are permitted for discussion.

The guard function accepts a single parameter input, representing the user input to your LLM application.

Interpreting Guard Result

print(guard_result.score)
print(guard_result.score_breakdown)

guard_result.score is an integer that is 1 if the guard has been breached. The score_breakdown for TopicalGuard a list of dictionaries, each containing:

• topic: The allowed topic against which the input was compared.
• score: A binary value (0 or 1), where 0 indicates the input aligns with the topic, and 1 indicates it does not.
• reason: An explanation of why the input does or does not align with the topic.

[
  {
    "topic": "technology",
    "score": 0,
    "reason": "The input discusses advancements in quantum computing, which falls under the technology topic."
  },
  {
    "topic": "science",
    "score": 0,
    "reason": "Quantum computing is a scientific field, aligning with the science topic."
  },
  {
    "topic": "health",
    "score": 1,
    "reason": "The input does not pertain to health-related matters."
  }
]

How Is it Calculated?

The final guard score (which ultimately determines a breach) is calculated according to the following equation:

$$\text{Topical Guard Score} = \begin{cases} 1 & \text{if no topic matches} \\ 0 & \text{if at least one topic matches} \end{cases}$$

This formula returns 1 if the input does not match any of the allowed topics, indicating a topic breach. It returns 0 if the input matches at least one allowed topic, indicating no breach.