wauldo-nemo 0.1.0

Wauldo fact-checking as a NeMo Guardrails output rail

wauldo-nemo

Wauldo fact-checking as a NeMo Guardrails output rail. Every bot message is verified against its retrieved context before it reaches the user; unsupported answers are refused, weakly-supported ones are annotated.

This is a thin adapter — all verdict logic lives in the wauldo SDK. This repo only maps a /v1/fact-check result to a rail decision and never re-implements parsing.

Install

pip install wauldo-nemo[nemo]

Requires wauldo>=0.19 (the version that exposes HttpClient.fact_check). While 0.19 is unreleased, install the SDK from source: pip install -e ../wauldo-monorepo/sdk-python.

How it works

bot message ──▶ wauldo_fact_check action ──▶ HttpClient.fact_check(text, context)
                                              │
                                              ▼
                                     policy.decide(result, thresholds)
                                              │
                          allow ──▶ pass   review ──▶ annotate   block ──▶ refuse

The policy never downgrades the server's verdict — thresholds can only make the rail stricter (e.g. strict=True turns review into a refusal, max_hallucination_rate forces a refusal above a ceiling).

Usage

from nemoguardrails import LLMRails, RailsConfig
from wauldo_nemo import register, PolicyThresholds

rails = LLMRails(RailsConfig.from_path("examples/config"))
register(rails, thresholds=PolicyThresholds(strict=True, max_hallucination_rate=0.4))

See examples/ for a runnable config (config.yml + rails.co).

Configuration

Threshold	Default	Effect
min_confidence	0.0	Below it, an allow is downgraded to annotate.
max_hallucination_rate	1.0	Above it, the response is refused.
strict	False	When set, a server review becomes a refusal.

Environment: WAULDO_API_KEY (required), WAULDO_BASE_URL (defaults to https://api.wauldo.com).

Develop

pip install -e '.[dev]'
pytest        # policy tests, no guardrails runtime needed

License

MIT