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
  • "Custom" detectors which can be defined with raw Python code, which allows for easy declaration of custom detection functions that can perform arbitrarily complex logic.

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.