gegentic/guardian

gegentic/guardian

Guardian is a lightweight, self-contained guardrail evaluation engine for LLM applications. It scores pro*** input and model output text against a set of configurable rules — PII detection, profanity, jailbreak atte***s, NSFW content, ban lists, and more — and returns a pass/fail decision with a reason.

This image is a public, educational build of the engine that powers gegentic.com. It runs fully on CPU with no external dependencies, making it easy to try out guardrail evaluation locally or in your own infrastructure.

📚 Full documentation, configuration reference, and usage guides are at docs.gegentic.com.

Quick start

bash
docker run -p 8700:8700 gegentic/guardian:latest

The service listens on port 8700 and exposes:

• GET /ping — health check
• POST /guardrails/evaluate — evaluate text against a policy set

Example request

bash
curl -X POST http://localhost:8700/guardrails/evaluate \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Some user-submitted text",
    "scope": "input",
    "rules": [{
      "nid": "rule-1",
      "action": "block",
      "scope": "both",
      "controls": [{
        "operator": "ANY",
        "criteria": [{
          "type": "profanity_check"
        }]
      }]
    }]
  }'

For the full request/response schema, the list of available guardrail types, and how to compose rules into policy sets, see docs.gegentic.com.

Tags

• latest, <version> — multi-arch manifest (amd64/arm64)
• latest-amd64, latest-arm64, <version>-amd64, <version>-arm64 — arch-specific images

Image notes

• Runs as a non-root user.
• CPU-only — no CUDA dependencies.
• Application code is compiled to native extensions; no .py source ships in the final image.
• pip/wheel are removed from the runtime image to reduce installer *** surface.

Support

Questions, issues, or feature requests: see docs.gegentic.com for support channels.