🤖AI Filter

The AI filter uses OpenAI's API to analyze chat messages for toxicity. It can work as a real-time filter or in post-hoc moderation mode where flagged messages are sent to the Moderation Queue for staff review.

Configuration

The AI filter is configured in moderation/ai-filter.yml:

enabled: false

# OpenAI API key (required)
api-key: ""

# AI model to use
model: "gpt-3.5-turbo"

# Toxicity score threshold (0.0 - 1.0)
# Messages scoring above this are flagged
toxicity-threshold: 0.7

# API request timeout (seconds)
timeout-seconds: 5

# Translation cache settings
cache-ttl-minutes: 30

# Rate limiting
max-requests-per-minute: 30

# Post-hoc moderation mode
post-hoc-moderation:
  enabled: false
  auto-delete-messages: false
  auto-learn-words: false

Modes

Real-Time Filtering

With post-hoc-moderation.enabled: false, messages are analyzed before being sent to chat. Messages exceeding the toxicity threshold are blocked.

Real-time filtering adds latency to every chat message due to the API call. Consider using post-hoc mode for better player experience.

Post-Hoc Moderation

With post-hoc-moderation.enabled: true, messages are sent to chat immediately and analyzed afterward. Flagged messages are sent to the Moderation Queue for staff review.

Setting	Description
auto-delete-messages	Automatically delete messages that exceed the threshold
auto-learn-words	Add frequently flagged words to the swear filter automatically

Key Settings

Setting	Default	Description
toxicity-threshold	0.7	Score threshold (lower = more strict)
max-requests-per-minute	30	API rate limit to control costs
cache-ttl-minutes	30	Cache identical messages to reduce API calls
model	gpt-3.5-turbo	OpenAI model to use

Requirements

• An OpenAI API key with billing enabled

• Network access from your Minecraft server to the OpenAI API