AI context check for ambiguous badwords

[raf-li]

Why

"anjing tetanggaku suka berisik kalau malam" is about a dog; "anjing lah" is a curse. The word list counts both, so students get warned for talking about their pets (or a tahi lalat, or mata telanjang).

What

• When every badword in a message is in AI_AMBIGUOUS_WORDS (animals, tahi, telanjang, seks, gay, ...), the bot asks an AI model on Cloudflare Workers AI whether it was an insult, in the background:
  • insult → counted as usual
  • ok → not counted
  • no answer (down, slow, busy) → not counted (agreed policy: better to miss a curse than warn someone about their dog)
• Clear badwords (kontol, bangsat, ...), or a message mixing one with an ambiguous word, are still counted immediately, without the AI.
• The bot never waits for the AI: the HTTP call runs on a background thread and results are applied on the main thread (the only one allowed to use the TeamTalk SDK).
• Conversation context: a lone "anjing" after "aku punya binatang baru waaa" is the pet, so the bot waits 8 s and also sends up to AI_CONTEXT_MESSAGES (default 4) earlier messages from the last 2 minutes and up to 2 follow-ups, labelled only [same person] / [someone else] (per channel: everyone in it; per PM: the sender only, since the SDK does not say who a PM was for). Kept in memory only while the check is on; commands and answers to the bot's prompts (passwords) are never kept. AI_CONTEXT_MESSAGES=0 sends the flagged message alone.
• Never nicknames, usernames or IPs. PMs included (agreed); mention it in the server rules.
• Off by default. /abt ai on switches it on and is refused until AI_CLOUDFLARE_ACCOUNT_ID and AI_CLOUDFLARE_API_TOKEN are in config.json. Merging this changes nothing on the server until then.
• /bwt <text> also shows the AI's verdict; tools/ai_eval.py scores the configured model on 38 labelled sentences (some with context) (dog vs curse, sate babi vs babi lu, mata telanjang vs harassment, ...).
• Default model @cf/meta/llama-3.3-70b-instruct-fp8-fast (configurable via AI_MODEL). Setup: docs/ai.md.

Tested

• Against a fake Workers AI server: request path (account, model), Authorization: Bearer, message layout, temperature: 0, answer parsing (JSON, extra text, bare word, errors), timeouts.
• In the bot with the real TeamTalk DLL: AI off → counted immediately as before; AI ok/insult/failure/busy → not counted / counted / not counted / not counted; clear and mixed badwords counted immediately; /bwt verdict PM; /abt ai on refused without credentials.
• Context: earlier message, someone else's question, follow-up, 4-message and 2-minute limits, PM sender only, passwords/commands never kept, nothing kept with the AI off or AI_CONTEXT_MESSAGES=0.
• All earlier suites still pass.
• Not yet run against the real Workers AI: run python3 tools/ai_eval.py on the server after adding the token.

🤖 Generated with Claude Code