integrations-haystack.md

Haystack Redaction Component

OpenMed provides a Haystack 2.x component that redacts Document.content locally before documents reach an index, retriever, prompt, or generator. Haystack remains optional: importing openmed or openmed.interop does not import it.

pip install "openmed[haystack]"

Redact documents before indexing

Place OpenMedRedactor before the first component that persists or embeds documents. The component returns copies, so the input documents are not modified.

from haystack import Document, Pipeline
from haystack.components.writers import DocumentWriter
from haystack.document_stores.in_memory import InMemoryDocumentStore

from openmed.interop.haystack import OpenMedRedactor

document_store = InMemoryDocumentStore()
indexing = Pipeline()
indexing.add_component("redactor", OpenMedRedactor())
indexing.add_component(
    "writer",
    DocumentWriter(document_store=document_store),
)
indexing.connect("redactor.documents", "writer.documents")

indexing.run(
    {
        "redactor": {
            "documents": [
                Document(
                    content=(
                        "Patient Jane Roe called jane.roe@example.com for follow-up."
                    ),
                    meta={"source": "synthetic-fixture"},
                )
            ]
        }
    }
)

Downstream query pipelines now retrieve the redacted content. Keep the redactor before splitting, embedding, caching, or writing so those stages never receive the original text. OpenMed performs de-identification on the local runtime; the component does not call a hosted Haystack service.

Configure de-identification

Masking and the deterministic safety sweep are enabled by default. Pass deidentify_kwargs to override arguments accepted by openmed.core.pii.deidentify().

redactor = OpenMedRedactor(
    deidentify_kwargs={
        "method": "mask",
        "confidence_threshold": 0.6,
        "lang": "en",
        "use_safety_sweep": True,
    }
)

Only Document.content is redacted. Treat document metadata, blobs, existing embeddings, and external caches as separate privacy surfaces and sanitize them before they enter the pipeline.