Groundedness — the HHEM backend

Output groundedness asks: is the model's answer supported by the source it was given, or did it make something up? Parapet ships two backends.

Lexical (default, zero-dependency)

The default backend is a dependency-free faithfulness proxy: it splits the answer into claims, checks each for support in the source, and treats a fabricated number (a wrong price or date) as a hard contradiction regardless of word overlap. It is honest about being a proxy — good enough to catch blatant invention, and it ships in the base install with nothing to set up.

from parapetai_agent.groundedness import score_groundedness
r = score_groundedness("The refund window is 90 days.", source="Refund window is 30 days.")
r.grounded        # False
r.contradiction   # True  (90 != 30)

HHEM (optional, higher-accuracy)

For production-grade faithfulness, enable the Vectara HHEM-2.1-Open cross-encoder (Apache-2.0) — a small model that scores factual consistency between a premise (the source) and a hypothesis (the answer).

It is not a declared extra: torch drags the whole CUDA wheel stack into a lockfile, heavy churn for a backend most deployments never enable. Install it explicitly:

pip install transformers torch

Then select the hhem backend in the groundedness config (a bundle entry, or score_groundedness(..., backend="hhem")). A bundle that names hhem without the libraries installed fails closed (the check errors and the caller denies) — it never silently turns groundedness off.

Running HHEM out-of-process

Loading transformers/torch into every agent process is heavy and duplicative. Point the agent at a shared, in-VPC HHEM scoring service instead:

export PARAPET_HHEM_URL="http://hhem.internal:8080"

The agent then POSTs {premise, hypothesis} pairs to {PARAPET_HHEM_URL}/score and reads back a consistency score — one model, many agents. Because HHEM is content-free and stateless (premise + hypothesis in, a float out), a single endpoint can serve every tenant safely. Swap between in-process and remote with that one env var; nothing else changes.

Content-free either way

Both backends run inside your process. Only the decision — grounded or not, and the score — becomes part of the (content-free) audit record. The source text and the answer never leave the process.