Chatstral

AI-powered chat moderation for Paper/Purpur servers. Every message is checked against a blacklist first, then against Shieldstral-1.0-3B, a local language model that judges messages against a plain-language moderation policy instead of a fixed keyword list.

Features

• Local LLM moderation: runs Shieldstral-1.0-3B through llama-server on your own machine, no chat text leaves the server
• Instant blacklist filter: plain words and regex patterns, checked before the AI runs
• Batched AI scoring: messages are batched and scored via logprob analysis on the model's yes/no answer
• Two filter modes: block cancels flagged messages, log lets everything through while still logging what would have been flagged, useful for tuning your threshold against real traffic
• Configurable policy: the instruct/query text the model judges messages against lives in config.yml, not hardcoded
• Auto-downloads llama-server and the model on first boot, no manual setup

Requirements

• Paper or Purpur, API version 26.1
• Java 25
• ~2GB free RAM for the model (CPU only, no GPU required)
• Outbound internet access on first boot to download llama-server and the model (a few GB combined)

Configuration

plugins/chatstral/config.yml:

filter-mode: block          # block or log
ai-filter-enabled: true
ai-threshold: 0.5
cooldown-ms: 1000
llama-port: 8734
llama-context-size: 2048
policy-instruct: "..."
policy-query: "..."

plugins/chatstral/blacklist.txt supports plain words and regex:-prefixed patterns.

Commands

- /chatstral (aliases /cs, /cfilter) — status
- /chatstral reload — reloads config and blacklist
- /chatstral status — model/download state

chatstral.bypass and chatstral.admin both default to op.