Local privacy proxy¶

OpenMed includes a small OpenAI-compatible chat-completions boundary for applications that already speak the common message-completion protocol. It listens locally, redacts text before calling an injected inference transport, and restores the placeholders only in the local response.

The proxy exposes both POST /v1/chat/completions and POST /chat/completions. A request uses the familiar shape:

{
  "model": "local-fixture",
  "messages": [
    {"role": "user", "content": "Contact Avery Example at 555-0100."}
  ]
}

Create the app with a local transport. The transport receives a copied request mapping whose text leaves contain deterministic OPENMED_PHI placeholders; the in-memory replacement map is not passed in metadata and is discarded when the request finishes.

from typing import Any

from openmed.service.privacy_proxy import create_app


def local_transport(payload: dict[str, Any], **_: Any) -> str:
    # Call an on-device model here. `payload` is already redacted.
    return payload["messages"][-1]["content"]


app = create_app(transport=local_transport)

Run that application with:

uvicorn my_proxy:app --host 127.0.0.1 --port 8081

The module-level openmed.service.privacy_proxy.app is import-safe, but has no transport configured and therefore rejects completion requests with a PHI-free 503 response. It never creates a network client. Inject an explicitly local transport before serving requests.

Set "stream": true to receive standard server-sent events. The transport may return text chunks, OpenAI-style chunk mappings, or a synchronous or asynchronous iterable of either. Placeholder tokens split across chunks are buffered until they can be restored; the replacement map itself is never sent as an event.

This is a privacy boundary, not a compliance certification or a clinical decision system. Use synthetic fixtures in development and validate residual risk for the deployment context.