Implementation:Protectai Llm guard Scan output

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

Overview

Concrete tool for sequentially scanning LLM outputs through a list of output scanners provided by the LLM Guard library.

Description

The scan_output function is the primary entry point for output validation in LLM Guard. It iterates over a list of OutputScanner instances, calling each scanner's scan method with the prompt and progressively sanitized output. It collects per-scanner validity flags and risk scores, and optionally stops early if fail_fast is enabled and a scanner fails.

Usage

Import this function when you need to validate and sanitize LLM-generated output before returning it to the user. This is the standard way to scan outputs in LLM Guard.

Code Reference

Source Location

Repository: llm-guard
File: llm_guard/evaluate.py
Lines: L76-127

Signature

def scan_output(
    scanners: list[OutputScanner],
    prompt: str,
    output: str,
    fail_fast: bool = False,
) -> tuple[str, dict[str, bool], dict[str, float]]:
    """
    Scans a given output of a large language model using the provided scanners.

    Args:
        scanners: A list of scanner objects inheriting from OutputScanner.
        prompt: The input prompt string that produced the output.
        output: The output string to be scanned.
        fail_fast: Stop scanning after the first scanner fails. Default False.

    Returns:
        A tuple containing:
            - The processed output string after applying all scanners.
            - A dictionary mapping scanner names to boolean validity flags.
            - A dictionary mapping scanner names to float risk scores (0=no risk, 1=high risk).
    """

Import

from llm_guard import scan_output

I/O Contract

Inputs

Name	Type	Required	Description
scanners	list[OutputScanner]	Yes	List of configured output scanner instances
prompt	str	Yes	The original prompt that produced the output
output	str	Yes	Raw LLM output string to scan
fail_fast	bool	No	Stop at first failing scanner (default: False)

Outputs

Name	Type	Description
sanitized_output	str	Cleaned output after all scanners applied
results_valid	dict[str, bool]	Scanner name to pass/fail mapping
results_score	dict[str, float]	Scanner name to risk score mapping (0.0-1.0)

Usage Examples

Basic Output Scanning

from llm_guard import scan_output
from llm_guard.output_scanners import Deanonymize, NoRefusal, Relevance, Sensitive
from llm_guard.vault import Vault

# 1. Configure output scanners
vault = Vault()  # Populated by prior Anonymize scan
output_scanners = [
    Deanonymize(vault),
    NoRefusal(threshold=0.75),
    Relevance(threshold=0.5),
    Sensitive(redact=True),
]

# 2. Scan the output
prompt = "What is the capital of France?"
llm_response = "The capital of France is Paris."
sanitized_output, results_valid, results_score = scan_output(
    output_scanners, prompt, llm_response
)

# 3. Check results
if all(results_valid.values()):
    print("Output is safe:", sanitized_output)
else:
    print("Output failed scanners:", {k: v for k, v in results_valid.items() if not v})

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment