Emet SDK for Python

Emet checks the claims an AI agent makes against the evidence behind them: the tool results and data the agent actually received. emet-sdk sends your agent's turns to Emet from the OpenTelemetry GenAI spans your framework already emits.

Early release: the API may change before 1.0.

uv add emet-sdk  # or: pip install emet-sdk

How it works

import emet_sdk
from pydantic_ai import Agent

emet_sdk.configure(service_name="support-agent", on_delivery=emet_sdk.write_delivery_file(".emet/last-delivery.json"))
Agent.instrument_all()


@app.post("/chat")
async def chat(body: ChatIn, user=Depends(current_user)):
    # Take these from your server's own auth and storage, never from the request body
    thread = await threads.get_for_user(user, body.thread_id)
    with emet_sdk.turn(tenant_ref=user.team_external_id, conversation_id=thread.id, user=user.email):
        result = await agent.run(body.message)
    return result.output

• Adds, never replaces. configure() adds one span processor to your app's TracerProvider. Logfire, Datadog or any other exporter keeps receiving the same spans, with the same sampling. Call it after your own tracing setup, such as logfire.configure(); it creates a provider only when the app has none.
• Opt in per turn. Only spans that start inside turn() are sent. Other agents and HTTP spans are not.
• Nothing written onto your spans. The tenant, conversation and user stay in the SDK and never reach your other exporters, or downstream services through baggage.
• No secrets, no system prompt. Secrets are redacted from everything sent, and the system prompt travels as a hash.
• The user is who triggered the turn, usually their email. Emet matches it to a member of your organization and stores it on the trace either way.
• Never in your way. Turns are sent on a background thread and never raise into your app. Without EMET_APPLICATION_KEY, the app logs "Emet is off" once and runs as before. Call emet_sdk.flush() before a short-lived process exits; Emet also flushes for up to 5 seconds at exit.
• Every turn reports back. on_delivery receives a result per turn: sent, stopped (with the reason, such as timeout or http_4xx and Emet's error code) or skipped, plus the trace, the span count and any tools Emet found missing arguments or results. write_delivery_file saves the latest one for emet-cli check --last; it never holds the key, the tenant, the user or any content.

Supported: Python 3.10 or later; Pydantic AI (Python 3.11 or later). Google ADK is planned.

Development

uv sync
uv run pytest
uv run python -m tests.record_fixtures && uv run pytest --update-fixtures  # rerecord, then review the diff
uv run ruff check . && uv run ruff format --check . && uv run pyright

To release, bump src/emet_sdk/_version.py and the changelog in a PR. Once it has merged, tag the merge commit vX.Y.Z and push the tag. The release workflow checks that the tag is on main and matches the version, builds and smoke-tests the wheel, then waits for approval in the release environment before publishing to PyPI through trusted publishing.

See CONTRIBUTING.md for how to contribute, and SECURITY.md to report a vulnerability.

License

Apache-2.0