Detector Algorithms for the FMS Guardrails Orchestrator

FMS Guardrails Orchestrator is an open source project led by IBM which provides a server for invocation of detectors on text generation input and output, and standalone detections.

This repository is intended to provide a collection of detector algorithms and microservices that are supported by the TrustyAI team.

Detectors

At the moment, the following detectors are supported:

• huggingface -- a generic detector class that is intended to be compatible with any AutoModelForSequenceClassification or a specific kind of AutoModelForCausalLM, namely GraniteForCausalLM; this detector exposes /api/v1/text/contents and thus, could be configured to be a detector of type: text_contents within the FMS Guardrails Orchestrator framework. This detector is also intended to be deployed as a KServe inference service.
• llm_judge -- Integrates the vLLM Judge library to use LLM-as-a-judge based guardrailing architecture
• builtIn -- Small, lightweight detection functions that are deployed out-of-the-box alongside the Guardrails Orchestrator. The built-in detectors provide a number of heuristic or algorithmic detection functions, such as:
  • Regex-based detections, with pre-written regexes for flagging various Personally Identifiable Information items like emails or phone numbers, as well as the ability to provide custom regexes
  • File-type validations, for verifying if model input/output is valid JSON, XML, or YAML

Building

To build the detector images, use the following commands:

Detector	Build Command
huggingface	podman build -t $TAG -f detectors/Dockerfile.hf detectors
llm_judge	podman build -t $TAG -f detectors/Dockerfile.judge detectors
builtIn	podman build -t $TAG -f detectors/Dockerfile.builtIn detectors

Replace $TAG with your desired image tag (e.g., my-detector:latest).

Running locally

Quick Start Commands

Detector	Run Command	Notes
builtIn	podman run -p 8080:8080 $BUILT_IN_IMAGE	Ready to use
huggingface	podman run -p 8000:8000 -e MODEL_DIR=/mnt/models/$MODEL_NAME -v $MODEL_PATH:/mnt/models/$MODEL_NAME:Z $HF_IMAGE	Requires model download
llm_judge	podman run -p 8000:8000 -e VLLM_BASE_URL=$LLM_SERVER_URL $LLM_JUDGE_IMAGE	Requires OpenAI-compatible LLM server

Detailed Setup Instructions & Examples

• Built-in detector: No additional setup required. Check out built-in detector examples to see how to use the built-in detectors for file type validation and personally identifiable information (PII) detection
• Hugging Face detector: Check out Hugging Face detector examples for a complete setup and examples on how to use the Hugging Face detectors for detecting toxic content and prompt injection
• LLM Judge detector: Check out LLM Judge detector examples for a complete setup and examples on how to use any OpenAI API compatible LLM for content assessment with built-in metrics and custom natural-language criteria

API

See IBM Detector API

License

This project is licensed under the Apache License Version 2.0 - see the LICENSE file for details.