TransformersExtractiveReader

Locate and extract answers to a query from documents using a Hugging Face question answering model. Use this component for extractive question answering pipelines where answers must come directly from source text.

Key Features

• Scores every possible answer span independently across documents for easier comparison.
• Supports long documents through configurable sequence length and stride-based splitting.
• Returns ranked ExtractedAnswer objects with scores and document offsets.
• Optionally returns a no-answer prediction when no span is confident enough.
• Removes duplicate overlapping answers using a configurable overlap threshold.

Configuration

1. Drag the TransformersExtractiveReader component onto the canvas from the Component Library.
2. Click on the component to open the configuration panel.
3. On the General tab:
   1. Set the model to a Hugging Face question answering model. The default is deepset/roberta-base-squad2-distilled.
   2. If the model requires authentication, create a secret with your Hugging Face API token. Use HF_API_TOKEN or HF_TOKEN as the environment variable name. For instructions, see Create Secrets.
4. Go to the Advanced tab to configure top_k, score_threshold, max_seq_length, stride, and no_answer.

note

Extractive readers run locally and may need GPU resources for larger models.

Connections

TransformersExtractiveReader receives a query string and a list of documents. Connect it after a retriever in a query pipeline. It outputs a list of ExtractedAnswer objects you can connect to AnswerBuilder or other downstream components.

Source Code

To check this component's source code, open extractive_reader.py in the Haystack Core Integrations repository.

Usage Examples

Basic Configuration

  TransformersExtractiveReader:
    type: haystack_integrations.components.readers.transformers.extractive_reader.TransformersExtractiveReader
    init_parameters:
      model: deepset/roberta-base-squad2-distilled
      top_k: 5
      no_answer: true

Using the Component in a Pipeline

# haystack-pipeline
components:
  retriever:
    type: haystack.components.retrievers.in_memory.embedding_retriever.InMemoryEmbeddingRetriever
    init_parameters:
      document_store:
        type: haystack.document_stores.in_memory.document_store.InMemoryDocumentStore
        init_parameters: {}
      top_k: 10

  reader:
    type: haystack_integrations.components.readers.transformers.extractive_reader.TransformersExtractiveReader
    init_parameters:
      model: deepset/roberta-base-squad2-distilled
      top_k: 3

connections:
- sender: retriever.documents
  receiver: reader.documents

max_runs_per_component: 100

metadata: {}

inputs:
  query:
  - reader.query

outputs:
  answers: reader.answers

Parameters

Inputs

Parameter	Type	Description
query	str	The question to answer.
documents	List[Document]	Documents to search for answer spans.
top_k	Optional[int]	The maximum number of answers to return. Overrides the init-time value.
score_threshold	Optional[float]	Returns only answers with a score above this threshold. Overrides the init-time value.
max_seq_length	Optional[int]	Maximum number of tokens before splitting a sequence. Overrides the init-time value.
stride	Optional[int]	Token overlap when a sequence is split. Overrides the init-time value.
max_batch_size	Optional[int]	Maximum number of samples processed at once. Overrides the init-time value.
answers_per_seq	Optional[int]	Number of answer candidates to consider per sequence. Overrides the init-time value.
no_answer	Optional[bool]	Whether to return a no-answer prediction. Overrides the init-time value.
overlap_threshold	Optional[float]	Removes duplicate answers when overlap exceeds this threshold. Overrides the init-time value.

Outputs

Parameter	Type	Description
answers	List[ExtractedAnswer]	Answers sorted by descending score. May include a no-answer entry when no_answer is enabled.

Init Parameters

These are the parameters you can configure in Pipeline Builder:

Parameter	Type	Default	Description
model	Union[str, Path]	deepset/roberta-base-squad2-distilled	A Hugging Face question answering model name or local path.
device	Optional[ComponentDevice]	None	The device on which the model is loaded. If None, automatically selected.
token	Optional[Secret]	Secret.from_env_var(["HF_API_TOKEN", "HF_TOKEN"], strict=False)	The API token used to download private models from Hugging Face.
top_k	int	20	Number of answers to return per query. Required even if score_threshold is set.
score_threshold	Optional[float]	None	Returns only answers with a probability score above this threshold.
max_seq_length	int	384	Maximum number of tokens before a sequence is split.
stride	int	128	Number of overlapping tokens when a sequence is split.
max_batch_size	Optional[int]	None	Maximum number of samples fed through the model at the same time.
answers_per_seq	Optional[int]	None	Number of answer candidates to consider per sequence. Defaults to 20 at run time.
no_answer	bool	True	Whether to return an additional no-answer prediction with an empty text and a score representing low confidence in the other answers.
calibration_factor	float	0.1	Factor used for calibrating probabilities.
overlap_threshold	Optional[float]	0.01	Removes duplicate answers when overlap exceeds this threshold. Set to None to keep all answers.
model_kwargs	Optional[Dict[str, Any]]	None	Additional keyword arguments passed to AutoModelForQuestionAnswering.from_pretrained.

Run Method Parameters

These are the parameters you can configure for the component's run() method. This means you can pass these parameters at query time through the API, in Playground, or when running a job. For details, see Modify Pipeline Parameters at Query Time.

Parameter	Type	Default	Description
query	str		The question to answer.
documents	List[Document]		Documents to search for answer spans.
top_k	Optional[int]	None	Maximum number of answers to return.
score_threshold	Optional[float]	None	Returns only answers above this score threshold.
max_seq_length	Optional[int]	None	Maximum tokens before splitting.
stride	Optional[int]	None	Token overlap when splitting.
max_batch_size	Optional[int]	None	Maximum batch size for inference.
answers_per_seq	Optional[int]	None	Answer candidates per sequence.
no_answer	Optional[bool]	None	Whether to include a no-answer prediction.
overlap_threshold	Optional[float]	None	Overlap threshold for deduplication.

Related Information

• Pipeline Components