Implementation:Protectai Llm guard Input BanCompetitors

Knowledge Sources	Protectai_Llm_guard
Domains	Content_Filtering, NER
Last Updated	2026-02-14 12:00 GMT

Overview

The BanCompetitors scanner detects and optionally redacts competitor company names in prompts using a Named Entity Recognition (NER) model.

Description

BanCompetitors is an input scanner that takes a user-provided list of competitor names and uses NER (Named Entity Recognition) to detect mentions of those competitors within prompts. It uses the guishe/nuner-v1_orgs model by default, which is specialized for organization entity recognition. When a competitor name is found, the scanner can either flag the prompt as invalid or redact the competitor name from the text. The scanner supports chunked processing with configurable chunk_size (default 512 tokens) and chunk_overlap_size (default 40 tokens) to handle long prompts efficiently. The threshold parameter controls the minimum NER confidence required for a match.

Usage

Use the BanCompetitors scanner when you need to prevent users from mentioning competitor organizations in their prompts. This is useful for brand protection, ensuring LLM responses do not reference or compare against competitor products, and for content moderation in customer-facing applications.

Code Reference

Source Location

• Repository: Protectai_Llm_guard
• File: llm_guard/input_scanners/ban_competitors.py
• Lines: 1-176

Signature

class BanCompetitors(Scanner):
    def __init__(
        self,
        competitors: Sequence[str],
        *,
        threshold: float = 0.5,
        redact: bool = True,
        model: Model | None = None,        # default: guishe/nuner-v1_orgs
        use_onnx: bool = False,
        chunk_size: int = 512,
        chunk_overlap_size: int = 40,
    ) -> None: ...

    def scan(self, prompt: str) -> tuple[str, bool, float]: ...

Import

from llm_guard.input_scanners import BanCompetitors

I/O Contract

Inputs

Name	Type	Required	Description
competitors	Sequence[str]	Yes	List of competitor names to detect in prompts.
threshold	float	No	Minimum NER confidence score for a match. Defaults to 0.5.
redact	bool	No	Whether to redact detected competitor names from the prompt. Defaults to True.
model	Model or None	No	The NER model to use. Defaults to guishe/nuner-v1_orgs.
use_onnx	bool	No	Whether to use ONNX runtime for inference. Defaults to False.
chunk_size	int	No	Token chunk size for processing long prompts. Defaults to 512.
chunk_overlap_size	int	No	Overlap between chunks to avoid missing entities at boundaries. Defaults to 40.

scan() Inputs

Name	Type	Required	Description
prompt	str	Yes	The input text to scan for competitor mentions.

Outputs

Name	Type	Description
prompt	str	The prompt with competitor names redacted (if redact=True), or the original prompt.
is_valid	bool	True if no competitor names were detected; False otherwise.
risk_score	float	A confidence score between 0.0 and 1.0 indicating the likelihood of competitor mention.

Usage Examples

Basic Usage

from llm_guard.input_scanners import BanCompetitors

scanner = BanCompetitors(
    competitors=["Google", "Microsoft", "Amazon"],
)
prompt = "How does our product compare to Google Cloud?"
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)

print(sanitized_prompt)  # Competitor name redacted
print(is_valid)          # False (competitor detected)
print(risk_score)        # Confidence score

Without Redaction

from llm_guard.input_scanners import BanCompetitors

# Detect but do not redact competitor names
scanner = BanCompetitors(
    competitors=["Apple", "Samsung"],
    redact=False,
    threshold=0.6,
)
prompt = "What features does Apple offer that we don't?"
sanitized_prompt, is_valid, risk_score = scanner.scan(prompt)

print(is_valid)  # False (competitor detected)
print(sanitized_prompt)  # Original prompt unchanged

Related Pages

• Principle:Protectai_Llm_guard_Competitor_Name_Detection