Implementation:Protectai Llm guard Relevance

Knowledge Sources	LLM Guard LLM Guard Documentation
Domains	NLP, Quality_Assurance, Semantic_Similarity
Last Updated	2026-02-14 12:00 GMT

Overview

Concrete tool for checking output relevance to prompts via embedding cosine similarity using BGE models, provided by the LLM Guard library.

Description

The Relevance class is an output scanner that encodes the prompt and output into dense vector embeddings using a BGE model (default: BAAI/bge-base-en-v1.5), then computes cosine similarity. Outputs with similarity below the threshold are flagged as irrelevant. Supports PyTorch and ONNX backends.

Usage

Import this scanner to verify LLM outputs are relevant to their prompts. Useful in pipelines where off-topic or hallucinated responses must be caught.

Code Reference

Source Location

Repository: llm-guard
File: llm_guard/output_scanners/relevance.py
Lines: L44-167

Signature

class Relevance(Scanner):
    def __init__(
        self,
        *,
        threshold: float = 0.5,
        model: Model | None = None,
        use_onnx: bool = False,
    ) -> None:
        """
        Args:
            threshold: Minimum similarity score. Default: 0.5.
            model: Embedding model. Default: BAAI/bge-base-en-v1.5.
            use_onnx: Use ONNX runtime. Default: False.
        """

    def scan(self, prompt: str, output: str) -> tuple[str, bool, float]:
        """
        Check output relevance to prompt via embedding similarity.

        Returns:
            - Original output (unmodified)
            - False if similarity below threshold, True otherwise
            - Risk score based on 1 - similarity
        """

Import

from llm_guard.output_scanners import Relevance

I/O Contract

Inputs

Name	Type	Required	Description
threshold	float	No	Minimum similarity score (default: 0.5)
model	Model	No	Embedding model (default: BAAI/bge-base-en-v1.5)
use_onnx	bool	No	Use ONNX runtime (default: False)
prompt	str	Yes (scan)	Original prompt
output	str	Yes (scan)	LLM output to check

Outputs

Name	Type	Description
output	str	Original output (unmodified)
is_valid	bool	False if similarity below threshold
risk_score	float	Risk score based on 1 - cosine similarity

Usage Examples

Basic Relevance Check

from llm_guard.output_scanners import Relevance

scanner = Relevance(threshold=0.5)

prompt = "What is the capital of France?"
relevant_output = "The capital of France is Paris."
_, is_valid, _ = scanner.scan(prompt, relevant_output)
# is_valid: True

irrelevant_output = "The best pizza recipe uses mozzarella cheese."
_, is_valid, score = scanner.scan(prompt, irrelevant_output)
# is_valid: False

Related Pages

Implements Principle

Principle:Protectai_Llm_guard_Output_Relevance_Checking

Requires Environment

Uses Heuristic

Heuristic:Protectai_Llm_guard_ONNX_Runtime_Optimization

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment