dragon.ai.inference.guardrails.GuardrailsProcessor

class GuardrailsProcessor

Bases: object

Handles prompt safety checking using PromptGuard model.

This class is responsible ONLY for guardrails/safety checking, completely separated from batching and LLM inference logic.

__init__(config: GuardrailsConfig, hf_token: str)

Initialize the guardrails processor.

Parameters:

• config (GuardrailsConfig) – Guardrails configuration.

• hf_token (str) – HuggingFace token for model access.

Methods

__init__(config, hf_token)	Initialize the guardrails processor.
check_prompts(prompts)	Check a list of prompts for jailbreak attempts.
filter_batch(prompts, formatted_prompts, ...)	Filter a batch of prompts, separating safe from malicious ones.
get_malicious_response()	Get the standard response for malicious prompts.

__init__(config: GuardrailsConfig, hf_token: str)

Initialize the guardrails processor.

Parameters:

• config (GuardrailsConfig) – Guardrails configuration.

• hf_token (str) – HuggingFace token for model access.

check_prompts(prompts: List[str]) → Tuple[List[bool], List[float], float]

Check a list of prompts for jailbreak attempts.

Parameters:

prompts (list[str]) – List of user prompts to check.

Returns:

Tuple (is_safe, jailbreak_scores, processing_time) where is_safe is a list of booleans (True if safe, False if malicious), jailbreak_scores are the scores per prompt, and processing_time is the total processing time in seconds.

Return type:

tuple[list[bool], list[float], float]

filter_batch(prompts: List[str], formatted_prompts: List[str], response_queues: List, latency_metrics: List[Tuple[float, float, float]]) → Tuple[List[str], List[str], List, List[Tuple[float, float, float]], List[int], float]

Filter a batch of prompts, separating safe from malicious ones.

Parameters:

• prompts (list[str]) – List of user prompts.

• formatted_prompts (list[str]) – List of formatted prompts.

• response_queues (list) – List of response queues.

• latency_metrics (list[tuple[float, float, float]]) – List of latency metric tuples.

Returns:

Tuple (safe_prompts, safe_formatted, safe_queues, safe_metrics, malicious_indices, processing_time) where the safe_* lists contain only safe entries, malicious_indices are the indices of malicious prompts and processing_time is the guardrails processing time in seconds.

Return type:

tuple

get_malicious_response() → str

Get the standard response for malicious prompts.

Returns:

Standard response string for malicious prompts.

Return type:

str