LLM, moderation & translation

Splintertree's LLM-driven features all read from a shared assignment table so a single provider switch (e.g. swap from Ollama Cloud to a local llama.cpp instance) updates every consumer.

Providers + models

• /admin/llm/providers (world.llm_provider) — base URL, API key, kind (openai / ollama). Supports any OpenAI-compatible endpoint.
• /admin/llm/models (world.llm_model) — per-provider model registry with name, context window, enabled flag.

Assignments

world.llm_assignment maps a use case to a model:

Use case	Default	Surface
moderation	glm-5.1	Forum + chat moderation pipeline
chat_translate	glm-5.1	In-game chat translation
npc_ai	glm-5.1	NPC dialogue agent
playbot_ai	glm-5.1	Player-bot decision agent
ticket_assist	glm-5.1	GM ticket reply drafter

Each assignment carries a system_prompt editable inline. The autonomous flag controls whether the agent acts without operator confirmation (default: false except for playbot_ai and npc_ai).

Moderation pipeline

• Backend: splintertree-common::moderation::classifier runs HMAC classifications via the configured LLM, caches verdicts per text+category in Redis.
• Categories (configurable in world.world_config.moderation_categories): spam, rmt, harassment, sexual, threats, personal_info.
• Threshold (moderation_threshold, default 70) flips Allow → Block.
• Forum traffic re-evaluates against a stricter local threshold (40) inside the web-api so a low-confidence harassment hit on a forum post still blocks even though the chat moderator caches it as Allow.
• Strikes (moderation_strike_threshold within moderation_strike_window_sec) trigger an automatic mute of moderation_mute_duration_sec.
• Chat / mail / forum surfaces are independently togglable (moderation_apply_to_chat, moderation_apply_to_mail).

Chat translate

• Live translation between chat locales when both sender and recipient are configured. The world worker batches translations per-locale per broadcast so a 35-player guild in mixed locales costs at most one LLM call per locale, not per recipient.
• Min character count (chat_translate_min_chars) avoids translating "lol" / "k" style noise.

Generative content (AdminAiGenPage)

Designer surface for generating NPC dialogue, quest text, or item descriptions from a brief — see Creatures & NPC AI for how that ties into the design pipeline.