Inference

The main entry point for hallucination detection.

HallucinationDetector(method='transformer', **kwargs)

Facade class that delegates to a concrete detector chosen by method.

Parameters:

Name	Type	Description	Default
method	str	"transformer" (token-classifier) or "llm" (OpenAI function-calling).	'transformer'
kwargs		Passed straight through to the chosen detector's constructor.	{}

Initialize the detector.

Parameters:

Name	Type	Description	Default
method	str	Detection method to use.	'transformer'
kwargs		Passed to the detector constructor.	{}

predict(context, answer, question=None, output_format='tokens', min_confidence=0.0)

Predict hallucination tokens or spans given passages and an answer.

This is the call most RAG pipelines use.

Parameters:

Name	Type	Description	Default
min_confidence	float	Drop "spans" whose confidence is below this threshold (in [0, 1]; 0.0 keeps every span). Ignored for "tokens" output. See the concrete detector docs for the structure of the returned list.	0.0

predict_prompt(prompt, answer, output_format='tokens', min_confidence=0.0)

Predict hallucinations when you already have a single full prompt string.

Parameters:

Name	Type	Description	Default
prompt	str	The prompt string.	required
answer	str	The answer string.	required
output_format	str	"tokens" to return token-level predictions, or "spans" to return grouped spans.	'tokens'
min_confidence	float	Drop "spans" below this confidence threshold ([0, 1]).	0.0

predict_prompt_batch(prompts, answers, output_format='tokens', min_confidence=0.0)

Batch version of :py:meth:predict_prompt.

Length of prompts and answers must match.

Parameters:

Name	Type	Description	Default
prompts	list[str]	List of prompt strings.	required
answers	list[str]	List of answer strings.	required
output_format	str	"tokens" to return token-level predictions, or "spans" to return grouped spans.	'tokens'
min_confidence	float	Drop "spans" below this confidence threshold ([0, 1]).	0.0