Skip to main content

04. PII across one conversation

pii(["EMAIL_ADDRESS"]) swaps an email address for a placeholder on the way in and restores it on the way out. One Redactions store per conversation keeps placeholders stable. pii is a local transform, so only the chat model differs between the two versions.

The email is caught by a local pattern, so this lesson needs no extra for pii. Saving the store with Redactions.dumps needs jes[crypto]. Only names (PERSON, on by default) need jes[pii] and a spaCy English model such as en_core_web_sm, which is why the lesson lists its one entity. See Multi-turn conversations.

Needs: OPENAI_API_KEY. No decision model runs: pii is local, and OpenAI writes the reply.

uv run --group examples python -m examples.04_pii.jev

View examples/04_pii/jev.py on GitHub

examples/04_pii/jev.py
"""Lesson 04 (Jev): Keep PII away from the model and give it back to the user.

pii() swaps an email address for a placeholder on the way in and restores it
on the way out. One Redactions store per conversation keeps placeholders
stable. pii is local, so only the chat model differs from local.py.

Run:   uv run --group examples python -m examples.04_pii.jev
Needs: OPENAI_API_KEY
"""

from langchain.chat_models import init_chat_model
from langchain_core.language_models.chat_models import BaseChatModel

from examples._common import print_check, require_env
from jes import Guard, Redactions
from jes.policies import pii


def chat_model() -> BaseChatModel:
    return init_chat_model("openai:gpt-5.4-mini", timeout=60, max_retries=2)


SYSTEM = (
    "Answer in one short sentence. Copy any [JES_PII_...] placeholder "
    "exactly as written; it stands for the user's real value."
)
# Demo only. Load a real 32-byte key from your secret manager.
STORE_KEY = b"k" * 32


def main() -> None:
    # The default entities include PERSON, which needs jes[pii] and a spaCy model.
    # Email is a local pattern, so this lesson runs without either.
    guard = Guard([pii(["EMAIL_ADDRESS"])])
    # The scope names the store; loading a saved store needs the same scope.
    store = Redactions(scope=b"conversation-1")

    incoming = guard.check_input(
        "Please confirm my email address: ada@example.com", redactions=store
    )
    reply = chat_model().invoke([("system", SYSTEM), ("user", incoming.onward)]).text
    outgoing = guard.check_output(reply, prompt=incoming, redactions=store)
    # Same store, so the same address gets the same placeholder.
    again = guard.check_input("ada@example.com again", redactions=store)

    # Persist the store encrypted between requests. associated_data binds the
    # blob to its context; loading it needs the same bytes.
    blob = store.dumps(STORE_KEY, associated_data=b"cookbook")

    print("== pii")
    print_check("input", incoming)
    print_check("output", outgoing)
    print_check("again", again)
    print(f"  reply: {outgoing.onward}")
    print(f"  saved store: {len(blob)} encrypted bytes")


if __name__ == "__main__":
    require_env("OPENAI_API_KEY")
    main()