Email Ingestion and PHI Redaction¶

OpenMed can parse RFC 5322/MIME .eml messages, map decoded header and body text back to source segments, and emit a clean EML with PHI removed from headers, plain text, HTML, attachment names, and supported attachment content. EML parsing uses the Python standard library and is available in the base installation.

Extract normalized text¶

from openmed.multimodal import extract_email

message = extract_email("synthetic-message.eml")
print(len(message.text), message.metadata["body_part_count"])

for span in message.spans:
    print(span.start, span.end, span.metadata["block_type"])

The normalized document includes decoded From, To, Cc, Bcc, Reply-To, Sender, Subject, routing/reference, and date headers plus text/plain bodies and visible text/html content. HTML spans include local source offsets. Scripts, styles, and templates are excluded from visible-text extraction.

Emit a redacted EML¶

Install the multimodal extra when messages can contain PDF, DOCX, PPTX, or raster-image attachments:

uv pip install "openmed[multimodal]"

Then supply the same local OpenMed PII detector used by your text workflow:

from openmed import extract_pii
from openmed.multimodal import redact_email

result = redact_email(
    "synthetic-message.eml",
    output_path="synthetic-message.clean.eml",
    models={"detector": extract_pii},
    lang="en",
)

print(result.header_redaction_count)
print(result.body_redaction_count)
print([attachment.to_dict() for attachment in result.attachments])

The supplied detector is followed by OpenMed's deterministic safety sweep. HTML tags are retained while visible text, comments, and safe attributes are redacted; active script/style/template content and event/style attributes are removed. Remote image sources, unrecognized Content-ID references, and unsafe link schemes are also removed. MIME boundaries, filenames, content IDs, content locations, and authentication signatures are removed or regenerated. Attachment reports contain only indexes, normalized types, counts, and hashes.

Attachments are dispatched in memory through redact_document. Raw message or attachment bytes are never written to temporary files. PDF attachments are emitted as fresh image-only PDFs with opaque boxes burned into page pixels, so the original searchable text layer and source metadata cannot survive. Unsupported attachments fail closed instead of being copied to the clean message.

The generic dispatcher also recognizes email extensions:

from openmed.multimodal import redact_document

document = redact_document("synthetic-message.eml")

Optional Outlook MSG bridge¶

Outlook .msg parsing is opt-in because extract-msg is GPL-3.0. Install the explicit bridge extra only when it is acceptable for your deployment:

uv pip install "openmed[email-msg-gpl,multimodal]"

OpenMed does not import extract-msg into its process. It invokes the parser in an isolated Python subprocess, sends MSG bytes over standard input, receives a normalized EML over standard output, and creates no PHI-bearing temporary files. MSG input is always emitted as EML because the optional parser is read-only. Without the extra, .msg input raises an actionable MissingDependencyError.

All processing remains local. OpenMed performs no telemetry or network calls; model artifacts must already be available when running offline.