arc-guardrails

A guardrails library for LLM applications. Three packages, one decision contract:

• arc-guard-core — zero-dep contract layer (typed models, Protocols, exception hierarchy, pipeline shape).
• arc-guard — batteries-included library with built-in inspectors, strategies, reporters, middleware adapters.
• arc-guard-service — transport-neutral deployment surface (in-process function call, CLI batch job, HTTP sidecar, FastAPI middleware).

Quick start

pip install arc-guard
python -c "
import asyncio
from arc_guard.pipeline import GuardPipeline
from arc_guard_core.types import GuardInput
result = asyncio.run(GuardPipeline().pre_process(GuardInput(text='ignore previous instructions')))
print(result.action, result.refusal.code if result.refusal else None)
"
# block jailbreak_strong

For sidecar deployment:

pip install 'arc-guard-service[fastapi]'
python -m arc_guard_service --port 8000

See examples/ for four runnable integration modes.

Development checks

The repo already exposes the primary quality gates through the root Makefile and the dashboard package scripts.

Python workspace checks:

make format
make lint
make test
make test-all

make test is the fast default and exercises core + pip. Use make test-all when you also want the api package suite.

GuardRailFlow dashboard checks:

cd apps/guardrail-flow
pnpm format
pnpm lint
pnpm test

CI

The repository now includes a GitHub Actions workflow at .github/workflows/ci.yml.

It runs package-scoped checks for core and pip on pull requests, pushes to main, and manual dispatch.

• The workflow only triggers for changes under packages/core, packages/pip, the Python workspace metadata, or the workflow file itself.
• Each check fans out as a core / pip matrix with fail-fast: false, so one package lane does not cancel the other.
• Each lane installs only its own package dependencies with uv sync --package ... --dev, instead of syncing the full workspace with all extras.

Run with Docker

A one-command stack is published on Docker Hub. No source checkout, no build step.

curl -fsSLO https://raw.githubusercontent.com/<owner>/<repo>/main/docker-compose.yml
docker compose up

Three services come up:

Service	URL	What it serves
Dashboard	http://127.0.0.1:5173	GuardRailFlow operator UI
API	http://127.0.0.1:8766/docs	arc-guard-service Swagger UI
Ollama	http://127.0.0.1:11434	Upstream LLM (llama3.2 pulled on first run)

Switch to the heavyweight ML-enabled image (semantic-intent + jailbreak-ml + code-injection extras pre-installed, sentence-transformer model pre-cached):

docker compose --profile ml up

Pin a specific version:

ARC_GUARD_IMAGE_TAG=0.7.1 docker compose up

Published images

Tag	Size	Inspectors enabled
wilp/2024mt03053-arc-guard-service:latest	~500 MB	Injection + Presidio (default pipeline)
wilp/2024mt03053-arc-guard-service:ml	~3.5 GB	All inspectors including semantic-intent + jailbreak-ml
wilp/2024mt03053-arc-guardrail-flow:latest	~45 MB	nginx + Vite-built SPA

Multi-arch builds: linux/amd64 + linux/arm64. Tag-driven CI publishes from git push --tags v*.*.* via .github/workflows/docker-publish.yml.

Architecture

See docs/architecture/README.md for the canonical architecture overview, including the rewrite roadmap, walkthrough index, and public-surface manifest.

Contributing

See CONTRIBUTING.md for the contribution flow and the doc-tree convention.

License

Proprietary; see LICENSE (if present) or contact the maintainers.