evennia/docs/source/Contribs/Contrib-Llm.md

Large Language Model ("Chat-bot AI") integration

Contribution by Griatch 2023

This adds an LLMClient that allows Evennia to send prompts to a LLM server (Large Language Model, along the lines of ChatGPT). Example uses a local OSS LLM install. Included is an NPC you can chat with using a new talk command. The NPC will respond using the AI responses from the LLM server. All calls are asynchronous, so if the LLM is slow, Evennia is not affected.

> create/drop villager:evennia.contrib.rpg.llm.LLMNPC
You create a new LLMNPC: villager

> talk villager Hello there friend, what's up?
You say (to villager): Hello there friend, what's up?
villager says (to You): Hello! Not much going on, really.

> talk villager Do you know where we are?
You say (to villager): Do you know where we are?
villager says (to You): We are in this strange place called 'Limbo'. Not much to do here.

Installation

You need two components for this contrib - Evennia, and an LLM webserver that operates and provides an API to an LLM AI model.

LLM Server

There are many LLM servers, but they can be pretty technical to install and set up. This contrib was tested with text-generation-webui. It has a lot of features while also being easy to install. |

1. Go to the Installation section and grab the 'one-click installer' for your OS.
2. Unzip the files in a folder somewhere on your hard drive (you don't have to put it next to your evennia stuff if you don't want to).
3. In a terminal/console, cd into the folder and execute the source file in whatever way it's done for your OS (like source start_linux.sh for Linux, or .\start_windows for Windows). This is an installer that will fetch and install everything in a conda virtual environment. When asked, make sure to select your GPU (NVIDIA/AMD etc) if you have one, otherwise use CPU.
4. Once all is loaded, stop the server with Ctrl-C (or Cmd-C) and open the file webui.py (it's one of the top files in the archive you unzipped). Find the text string CMD_FLAGS = '' near the top and change this to CMD_FLAGS = '--api'. Then save and close. This makes the server activate its api automatically.
5. Now just run that server starting script (start_linux.sh etc) again. This is what you'll use to start the LLM server henceforth.
6. Once the server is running, point your browser to http://127.0.0.1:7860 to see the running Text generation web ui running. If you turned on the API, you'll find it's now active on port 5000. This should not collide with default Evennia ports unless you changed something.
7. At this point you have the server and API, but it's not actually running any Large-Language-Model (LLM) yet. In the web ui, go to the models tab and enter a github-style path in the Download custom model or LoRA field. To test so things work, enter DeepPavlov/bart-base-en-persona-chat and download. This is a relatively small model (350 million parameters) so should be possible to run on most machines using only CPU. Update the models in the drop-down on the left and select it, then load it with the Transformers loader. It should load pretty quickly. If you want to load this every time, you can select the Autoload the model checkbox; otherwise you'll need to select and load the model every time you start the LLM server.
8. To experiment, you can find thousands of other open-source text-generation LLM models on huggingface.co/models. Beware to not download a too huge model; your machine may not be able to load it! If you try large models, don't set the Autoload the model checkbox, in case the model crashes your server on startup.

For troubleshooting, you can look at the terminal output of the text-generation-webui server; it will show you the requests you do to it and also list any errors. See the text-generation-webui homepage for more details.

Evennia config

To be able to talk to NPCs, import and add the evennia.contrib.rpg.llm.llm_npc.CmdLLMTalk command to your Character cmdset in mygame/commands/default_commands.py (see the basic tutorials if you are unsure).

The default LLM api config should work with the text-generation-webui LLM server running its API on port 5000. You can also customize it via settings (if a setting is not added, the default below is used):

    # path to the LLM server
    LLM_HOST = "http://127.0.0.1:5000"
    LLM_PATH = "/api/v1/generate"

    # if you wanted to authenticated to some external service, you could
    # add an Authenticate header here with a token
    LLM_HEADERS = {"Content-Type": "application/json"}

    # this key will be inserted in the request, with your user-input
    LLM_PROMPT_KEYNAME = "prompt"

    # defaults are set up for text-generation-webui and most models
    LLM_REQUEST_BODY = {
        "max_new_tokens": 250,  # set how many tokens are part of a response
        "temperature": 0.7, # 0-2. higher=more random, lower=predictable
    }
    # helps guide the NPC AI. See the LLNPC section.
    LLM_PROMPT_PREFIx = (
      "You are roleplaying as {name}, a {desc} existing in {location}. "
      "Answer with short sentences. Only respond as {name} would. "
      "From here on, the conversation between {name} and {character} begins."
 ​)

Don't forget to reload Evennia if you make any changes.

Usage

With the LLM server running and the new talk command added, create a new LLM-connected NPC and talk to it in-game.

> create/drop girl:evennia.contrib.rpg.llm.LLMNPC
> talk girl Hello!
You say (to girl): Hello
girl ponders ...
girl says (to You): Hello! How are you?

Most likely, your first response will not be this nice and short, but will be quite nonsensical, looking like an email. This is because the example model we loaded is not optimized for conversations. But at least you know it works!

The conversation will be echoed to everyone in the room. The NPC will show a thinking/pondering message if the server responds slower than 2 seconds (by default).

A note on running LLMs locally

Running an LLM locally can be very demanding.

As an example, I tested this on my very beefy work laptop. It has 32GB or RAM, but no gpu. so i ran the example (small 128m parameter) model on cpu. it takes about 3-4 seconds to generate a (frankly very bad) response. so keep that in mind.

On huggingface.co you can find listings of the 'best performing' language models right now. This changes all the time. The leading models require 100+ GB RAM. And while it's possible to run on a CPU, ideally you should have a large graphics card (GPU) with a lot of VRAM too.

So most likely you'll have to settle on something smaller. Experimenting with different models and also tweaking the prompt is needed.

Also be aware that many open-source models are intended for AI research and licensed for non-commercial use only. So be careful if you want to use this in a commercial game. No doubt there will be a lot of changes in this area over the coming years.

Why not use an AI cloud service?

You could in principle use this to call out to an external API, like OpenAI (chat-GPT) or Google. Most cloud-hosted services are commercial and costs money. But since they have the hardware to run bigger models (or their own, proprietary models), they may give better and faster results.

Calling an external API is not tested, so report any findings. Since the Evennia Server (not the Portal) is doing the calling, you are recommended to put a proxy between you and the internet if you call out like this.

Here is an untested example of the Evennia setting for calling OpenAI's v1/completions API:

LLM_HOST = "https://api.openai.com"
LLM_PATH = "/v1/completions"
LLM_HEADERS = {"Content-Type": "application/json", 
               "Authorization": "Bearer YOUR_OPENAI_API_KEY"}
LLM_PROMPT_KEYNAME = "prompt"
LLM_REQUEST_BODY = { 
                        "model": "gpt-3.5-turbo",
                        "temperature": 0.7,
                        "max_tokens": 128,
                   }

TODO: OpenAI's more modern v1/chat/completions api does currently not work out of the gate since it's a bit more complex, having the prompt given as a list of all responses so far.

The LLMNPC class

The LLM-able NPC class has a new method at_talked_to which does the connection to the LLM server and responds. This is called by the new talk command. Note that all these calls are asynchronous, meaning a slow response will not block Evennia.

The NPC's AI is controlled with a few extra properties and Attributes, most of which can be customized directly in-game by a builder.

prompt_prefix

The prompt_prefix is very important. This will be added in front of your prompt and helps the AI know how to respond. Remember that an LLM model is basically an auto-complete mechaniss, so by providing examples and instructions in the prefix, you can help it respond in a better way.

The prefix string to use for a given NPC is looked up from one of these locations, in order:

1. An Attribute npc.db.chat_prefix stored on the NPC (not set by default)

2. A property chat_prefix on the the LLMNPC class (set to None by default).

3. The LLM_PROMPT_PREFIX setting (unset by default)

4. If none of the above locations are set, the following default is used:

   "You are roleplaying as {name}, a {desc} existing in {location}.
   Answer with short sentences. Only respond as {name} would.
   From here on, the conversation between {name} and {character} begins."
   

Here, the formatting tag {name} is replaced with the NPCs's name, desc by it's description, the location by its current location's name and character by the one talking to it. All names of characters are given by the get_display_name(looker) call, so this may be different from person to person.

Depending on the model, it can be very important to extend the prefix both with more information about the character as well as communication examples. A lot of tweaking may be necessary before producing something remniscent of human speech.

Response template

The response_template AttributeProperty defaults to being

$You() $conj(say) (to $You(character)): {response}"

following common msg_contents FuncParser syntax. The character string will be mapped to the one talking to the NPC and the response will be what is said by the NPC.

Memory

The NPC remembers what has been said to it by each player. This memory will be included with the prompt to the LLM and helps it understand the context of the conversation. The length of this memory is given by the max_chat_memory_size AttributeProperty. Default is 25 messages. Once the memory is maximum is reached, older messages are forgotten. Memory is stored separately for each player talking to the NPC.

Thinking

If the LLM server is slow to respond, the NPC will echo a random 'thinking message' to show it has not forgotten about you (something like "The villager ponders your words ...").

They are controlled by two AttributeProperties:

• thinking_timeout: How long, in seconds to wait before showing the message. Default is 2 seconds.
• thinking_messages: A list of messages to randomly pick between. Each message string can contain {name}, which will be replaced by the NPCs name.

TODO

There is a lot of expansion potential with this contrib. Some ideas:

• Easier support for different cloud LLM provider API structures.
• More examples of useful prompts and suitable models for MUD use.

---

This document page is generated from evennia/contrib/rpg/llm/README.md. Changes to this file will be overwritten, so edit that file rather than this one.