panda_guard.pipelines package

Submodules

panda_guard.pipelines.inference module

class panda_guard.pipelines.inference.InferPipeline(config: InferPipelineConfig, verbose: bool = False)

Bases: object

Inference pipeline for managing and executing attacks and defenses.

Parameters:

• config – Configuration for the inference pipeline, including attacker, defender, and judges.

• verbose – Whether to enable verbose logging for debugging and transparency.

calc_tokens() → Dict[str, Dict[str, int]]

Calculate the token usage for both the attacker and defender, summarizing prompt and completion tokens.

Returns:

Dictionary containing detailed token usage for both roles.

log(messages: List[Dict[str, str]], stage: str | None = None)

Log messages for each stage of the pipeline if verbose logging is enabled.

Parameters:

• messages – The messages to be logged at the current stage.

• stage – A label for the current stage of the pipeline.

parallel_judging(defense: List[Dict[str, str]], request: str) → Dict[str, Any]

Perform parallel evaluation of the defense using multiple judges.

Parameters:

• defense – Messages that have been defended and need evaluation.

• request – The original request for context in evaluation.

Returns:

A dictionary containing the results from each judge.

reset()

Reset the state of all LLMs in the pipeline, preparing them for a new inference run.

class panda_guard.pipelines.inference.InferPipelineConfig(attacker_config: ~panda_guard.role.attacks.base.BaseAttackerConfig | None = None, defender_config: ~panda_guard.role.defenses.base.BaseDefenderConfig | None = None, judge_configs: ~typing.List[~panda_guard.role.judges.base.BaseJudgeConfig] = <factory>)

Bases: object

Configuration class for the inference pipeline.

Parameters:

• attacker_config – Configuration for the attacker, detailing its parameters and behavior.

• defender_config – Configuration for the defender, detailing its parameters and behavior.

• judge_configs – List of configurations for any judges used in evaluating the effectiveness of the defense.

attacker_config: BaseAttackerConfig = None

defender_config: BaseDefenderConfig = None

judge_configs: List[BaseJudgeConfig]

panda_guard.pipelines.inference.llm_register(attr: BaseAttacker | BaseDefender | BaseJudge) → List[BaseLLM]

Register any LLM instances associated with the given attribute (attacker, defender, or judge).

Parameters:

attr – An instance of an attacker, defender, or judge.

Returns:

List of registered LLM instances.

Module contents