Skip to main content
For the complete documentation index for agents and LLMs, see llms.txt.

PresidioDocumentCleaner

Remove personally identifiable information (PII) from Haystack Document objects using Microsoft Presidio, replacing detected entities with type placeholders such as <PERSON> or <EMAIL_ADDRESS>.

Key Features​

  • Detects and anonymizes PII in document text content using Presidio's analyzer and anonymizer engines.
  • Replaces detected entities with labeled placeholders, for example <PERSON> and <PHONE_NUMBER>, preserving document structure.
  • Runs locally — no external API call or API key required.
  • Supports multiple languages; automatically selects the appropriate spaCy model for built-in languages such as English, German, French, and Spanish.
  • Documents without text content pass through unchanged.
  • Configurable confidence threshold and entity type allowlist for fine-grained control over what gets anonymized.

Configuration​

  1. Drag the PresidioDocumentCleaner component onto the canvas from the Component Library.
  2. Click on the component to open the configuration panel.
  3. On the General tab:
    • Set language to the ISO 639-1 code of your document language (for example, en for English).
    • Optionally, set entities to restrict anonymization to specific PII types.
  4. Go to the Advanced tab to adjust score_threshold or specify custom models for unsupported languages.
spaCy model dependency

PresidioDocumentCleaner requires a spaCy language model. For English, install en_core_web_lg with python -m spacy download en_core_web_lg. For other built-in languages, install the corresponding model listed in the spaCy model documentation. The component loads the model automatically on the first run() call.

Connections​

PresidioDocumentCleaner receives a list of documents, typically from a document converter or document store retriever. It outputs a new list of documents with PII removed. Connect its documents output to downstream components such as PromptBuilder or a document writer.

Source Code​

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

Usage Examples​

Basic Configuration​

  PresidioDocumentCleaner:
    type: haystack_integrations.components.preprocessors.presidio.presidio_document_cleaner.PresidioDocumentCleaner
    init_parameters:
      language: en
      entities:
      score_threshold: 0.35
      models:

Using the Component in a Pipeline​

# haystack-pipeline
components:
  PresidioDocumentCleaner:
    type: haystack_integrations.components.preprocessors.presidio.presidio_document_cleaner.PresidioDocumentCleaner
    init_parameters:
      language: en
      entities:
        - PERSON
        - EMAIL_ADDRESS
        - PHONE_NUMBER
      score_threshold: 0.35
      models:

connections: []

max_runs_per_component: 100

metadata: {}

inputs:
  documents:
  - PresidioDocumentCleaner.documents

outputs:
  documents: PresidioDocumentCleaner.documents

Parameters​

Inputs​

ParameterTypeDescription
documentsList[Document]Documents whose text content will be anonymized.

Outputs​

ParameterTypeDescription
documentsList[Document]Cleaned documents with PII replaced by entity type placeholders. Documents without text content are passed through unchanged.

Init Parameters​

These are the parameters you can configure in Pipeline Builder:

ParameterTypeDefaultDescription
languagestr"en"ISO 639-1 language code for PII detection. For built-in languages such as "de", "fr", and "es", the appropriate spaCy model is loaded automatically. For unsupported languages, use models to specify a custom model. See Presidio supported languages.
entitiesOptional[List[str]]NoneList of PII entity types to detect and anonymize, for example ["PERSON", "EMAIL_ADDRESS"]. If None, all supported entity types are used. See Presidio supported entities.
score_thresholdfloat0.35Minimum confidence score (0–1) for a detected entity to be anonymized.
modelsOptional[List[Dict[str, str]]]NoneAdvanced override: list of spaCy model configurations. Each entry must contain "lang_code" and "model_name" keys, for example [{"lang_code": "fr", "model_name": "fr_core_news_md"}]. Use this only when you need a specific model variant or a language not covered by the built-in mapping.