Files
moonwell-core/modules/mod-playerbots-ai/README.md
T

115 lines
5.0 KiB
Markdown

# MoonWell Playerbots AI
This module sends a local `/say` or public-channel message to the LLM only when
a real player mentions the complete name of an online bot. A channel bot must
be a channel member; a `/say` bot must be within the server's normal listening
range. HTTP work runs outside the AzerothCore world thread.
Alice AI LLM Flash can classify direct natural-language requests into a small
allowlisted set of native playerbots commands. The server validates the command
again and executes it only against the addressed bot; existing playerbots
ownership and group security checks still apply.
The routing prompt covers party movement and combat, LFG, combat/non-combat
strategies, spells, RTI/RTSC, focus healing, loot lists, inventory operations,
quests, pets, talents, glyphs, maintenance, and information commands documented
in the upstream Playerbot Commands wiki. Administrative, console, account,
debug/cheat, and destructive `destroy`/`drop`/mail commands are intentionally
not executable through the LLM.
Common Russian movement/combat phrases use a deterministic fast path. Other
requests first pass through a command-only classifier; ordinary conversation is
then sent separately to the roleplay prompt. While AI routing is enabled, the
legacy public-channel parser is bypassed and `setup-modules.sh` also forces
`AiPlayerbot.RandomBotTalk = 0`. RandomBotTalk consumes incoming chat packets
inside each bot, so bypassing only the parser does not prevent its delayed
stock replies. Whisper, party, raid, and guild command hooks remain unchanged.
Recognized commands produce a gender-aware local-chat acknowledgement
(`Понял команду.` or `Поняла команду.`) before they are handed to the native
playerbots command handler. Bot and player gender are also included in the LLM
context for pronoun, past-tense verb, and adjective agreement. Russian requests
to show bags or inventory use the native `inv` command; `open items` is reserved
for actually opening container items.
Trade and quest requests have deterministic link-preserving routes before the
LLM classifier:
- `дай/передай [item]` -> `t [item]`; `начни обмен` -> `t`
- `покажи задания` -> `quests all`; completed/incomplete filters are supported
- `покажи прогресс [quest]` -> `q [quest]`
- `прими [quest]` / `прими все задания` -> `accept [quest]` / `accept *`
- `поделись [quest]` -> `share [quest]`
- `выбери [item] в награду` -> `r [item]`
- `сдай готовое задание` -> `talk`
- `откажись от [quest]` -> `drop [quest]` only with an explicit quest link
These routes preserve the original WoW hyperlink byte-for-byte. The local
playerbots trade action continues processing after it creates the trade, so one
`t [item]` command both initiates the exchange and places the requested item in
the trade window.
The bundled `playerbots-ai` bridge uses the Yandex AI Studio OpenAI-compatible
Chat Completions API and defaults to Alice AI LLM Flash.
## Configuration
Set these values in the repository `.env`:
```dotenv
ACORE_PLAYERBOTS_AI_ENABLED=1
ACORE_PLAYERBOTS_AI_COMMANDS_ENABLED=1
YANDEX_AI_API_KEY=<service-account API key>
YANDEX_AI_FOLDER_ID=<Yandex Cloud folder ID>
```
Until the subscription system exists, public access is controlled by one
explicit switch:
```dotenv
ACORE_PLAYERBOTS_AI_AVAILABLE_FOR_ALL=1
```
`1` allows every player to reach the bridge. `0` denies LLM access to everyone
without spending tokens. When subscriptions are implemented, their entitlement
check will be added to the disabled branch instead of using a temporary manual
account list.
The command classifier has a persistent SQLite cache in the
`playerbots-ai-cache` Docker volume. The key excludes the addressed bot name,
so repeated phrases such as “покажи инвентарь” are shared across players and
bots. Both commands and “ordinary chat, not a command” decisions are cached.
The cache does not store roleplay responses or conversation history.
Obvious social phrases such as greetings, “как дела?” and “расскажи о себе”
skip the command classifier entirely: they still receive an Alice roleplay
reply, but do not pay the roughly 800-token command-catalog prompt first.
```dotenv
PLAYERBOTS_AI_CLASSIFIER_CACHE_ENABLED=1
PLAYERBOTS_AI_CLASSIFIER_CACHE_TTL_SECONDS=604800
PLAYERBOTS_AI_CLASSIFIER_CACHE_MAX_ENTRIES=10000
```
Changing the classifier prompt automatically creates a new cache namespace.
Old rows expire normally, and the oldest rows are removed at the size limit.
The default model URI is:
```text
gpt://<YANDEX_AI_FOLDER_ID>/aliceai-llm-flash
```
Run `./start-server.sh` or `scripts/prod-deploy.sh`. The deployment wrapper
builds and starts the bridge before `ac-worldserver`.
Never commit a populated API key.
Full Yandex response JSON is logged by default by the bridge:
```dotenv
PLAYERBOTS_AI_LOG_API_RESPONSES=1
```
Use `docker compose logs -f playerbots-ai` to follow classifier and chat
responses. Authentication headers and the API key are never logged.