MONITAUR

This is the main repository of the Monitaur project, an open-source platform for monitoring AI services. Funded by Netidee.

Monitoring API with Prometheus

This is a lightweight, extensible monitoring server built with FastAPI, integrated with Prometheus for metrics collection. It leverages the reusable monitoring_toolkit package to detect suspicious model inputs (e.g. low-confidence, repetitive, or anomalous queries).

Features

• Plug-and-play detector system (monitoring_toolkit)
• Built-in Prometheus metrics (suspicious counts, confidence scores, errors, etc.)
• FastAPI server to expose detection as an HTTP API
• Docker support for full stack orchestration

Installation

Clone the repo and start the full monitoring stack:

git clone https://github.com/sbaresearch/monitaur.git
cd monitaur
docker compose up --build

The API will be running at http://localhost:8000, and Prometheus metrics at http://localhost:9090.

API Usage

Send detection requests to the /detect endpoint:

POST /detect
Content-Type: application/json

{
  "input_data": "This is a suspicious input",
  "model_output": [0.25, 0.25, 0.25, 0.25],
  "modality": "text"
}

You’ll get a response like:

{
  "is_suspicious": true,
  "confidence": 1.38,
  "reason": "entropy mode: suspicious score = 1.3863, threshold = 0.8000"
}

Quick Testing with client.py

You can also test the detection endpoint using the built-in client.py script:

python app/client.py

It will send a few example requests and print the detection results to the console.

Prometheus Metrics

Metrics are exposed at /metrics. You’ll find:

• detector_requests_total: All incoming detection queries
• suspicious_queries_total: Total flagged queries
• detection_confidence_score: Histogram of confidence scores
• detector_errors_total: Internal processing errors

These metrics can be scraped by Prometheus and visualized (e.g. via Grafana).

Metrics can be extended with additional fields from each DetectionResult, and customized further depending on the active detector’s behavior (e.g., tracking hash collisions for repetition, or distances for similarity).

Customization

To switch or extend detectors:

• Modify parameters of get_detector in app/detectors.py
• Update metrics.py accordingly to DetectionResult of the new detector
• All detectors from monitoring_toolkit are supported:
  • confidence, repetition, similarity
  • Modal-specific versions: text_similarity, image_repetition, etc.

Want to add your own? See the toolkit's README for instructions on writing and registering new detectors.

Docker Compose Setup

Here's what's inside docker-compose.yml:

• api: Runs the FastAPI server
• prometheus: Collects metrics

Requirements

• Python 3.10+
• Docker and Docker Compose
• Install Python dependencies via:

  pip install -r requirements.txt