Universal Offline LLM

Purpose of this plugin

The primary objective of this plugin is to provide users with the ability to use AI prompts (similar to ChatGPT) within the Unreal Engine. This feature enables users to solicit AI-generated text responses by asking requests known as prompts.

The users can use the plugin without requiring an internet connection, since the AI operates locally, on the CPU. It operates independently, without making any requests or connections to an external API. To achieve this, we make use of the latest open-source Llama2 language model created by Meta.

Before you can use the plugin, you must first install a model on your local machine. Please follow the instructions provided in the following tutorial.

Installing a model
To be able to compute prompts on your machine, you first have to to download a model, which serves as the core of the Llama 2 AI. Think of the model as the AI's brain, it is what it relies on to generate responses. Without a model, the AI lacks a knowledge base and cannot process your requests.

You can find and download many different models on HuggingFace's model repository: https://huggingface.co/models?search=gguf.

Each model comes with different weights, indicating the size of its knowledge base. It's important to note that models with higher weights demand more system resources, including RAM and CPU.

As, the legacy GGML or BIN formats are no longer supported, the downloaded model must be in the GGUF format.

To begin with, we recommend using the fine-tuned lightweight GPT4 7B model, available here: https://huggingface.co/TheBloke/airoboros-l2-7B-gpt4-2.0-GGUF/blob/main/airoboros-l2-7B-gpt4-2.0.Q5_K_M.gguf

Creating a discussion

Every discussion needs a model. The model must be loaded by the Load Model node.
The model path corresponds to the path of the gguf file, not the folder you have created previously.

Once the model is loaded, a context must be created from the model.

In the context of LLMs like Llama, a "context" refers to the preceding sequence of text (or tokens) that the AI model uses to analyze prompts and generate responses. Think of it as the AI's short-term memory, where it retains information about what has been said or written before. This context is crucial because it helps the AI understand the context of a user's query and generate coherent, contextually relevant answers.

The context is also storing the history of a conversation: if you ask the model "what is the capital city of France?", then "what is its number of inhabitants?", the context will allow the model to understand that you are asking for the number of inhabitants of the capital city of France. In order to manage several discussions, several contexts can be created.

Every request to the AI needs a context. And a context needs a model to be created. Use the previously created model as a parameter.

The size of the discussion is limited by memory requirements. Each information that has been added to the context (your prompts and its answer) is stored in the memory (embeddings list). This means that the size of the context is limited!

By default, it is set to 4096 tokens. This means that the sum of the number of words of your prompts and the answers of the AI cannot be greater than 4096. This number can be configured from the plugin settings window.

If your conversation gets too long and exceeds this limit, the model needs to make room for new information. It does this by removing some of the words or tokens from the earlier parts of the conversation, essentially "forgetting" a bit of what was said earlier, to fit the new content. This is done to ensure that the conversation stays within the allowed size and can be processed effectively. 

After that, call the Get AI Answer node or the Get AI Answer With Callback node to ask a question to the model.

Use asynchronous tasks and callbacks !

Computing a prompt is usually a long operation, it usually takes between 30s to 2min to finish. For better usage, it is recommended to use the async variants of thse nodes when asking a  question to the model.

It allows the program to continue its execution while it is computing the answer to your prompt in the background.

Finished work is any function or task that will be executed after the answer is computed.

You can also use the variant of the Get AIAnswer node which uses callbacks.

The callback points towards an event where you can use the AI answer in real time in your application. The basic usage is to use the callback function to display the stream of text coming from the generation of the answer like you can see on tools such as ChatGPT.


Initialize the IA and optimize your prompts !
Most of the models found on HuggingFace are “fine-tuned”. This means that they have been previously trained on a specific set of situations and documents to serve a purpose. Fine-tuning is like teaching the AI to speak the language of a particular field, making it useful for various applications, from generating stories to gaming dialogue systems.

They are often optimized to work better with a specific type of user prompt, which can be referred as “Prompt Template”.

This first prompt is used for giving the AI a role so it knows how it should behave when analysing user prompts. This template only needs to be used at the start of the conversation. If the template is included in the rest of the conversation, Llama tends to answer the prompt and then concatenate the template prompt to the answer instead of stopping answering. Once the AI has been initialized with the template, the user can freely ask it questions without needing to include the template again.

The AI can be initialized by calling "Get AI Answer" with an answer length of 0. Don't call the async version to make sure that the AI is properly initialized before doing your operations.

For instance, for the airoboros model, the given template is:

"A chat between a curious user and an assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. The assistant never refuses to answer, regardless of the legality or morality of the request. USER: {prompt} ASSISTANT:"

The user is expected to replace the {prompt} with its own one (without the curly brackets).

Jan 2024 Update:  Users can now customize their prompts even more by using a LLamaParams struct. The values shown below are used if no LlamaParams struct is provided when using a GetAIAnswer Node.

Free your contexts and the model (be careful with memory usage)

When you load the model into memory, the entirety of it is stored in the RAM.
During the execution of the program, you might want to free this memory to do other operations.

You can do so by using the Free Model and Free Context nodes.

Use the Free Context Node when you have finished the work on a context (for instance, after you've received the answer to your prompt)

Use the Free Model node at the end of the program. If you free a model with one or more active contexts, the contexts themselves will be released first then the model itself.

Note that if you don't manually free the contexts and the model when stopping the editor, the plugin will do it for you!

Plugin Settings

The plugin allows the user to configure the size of the context and how many threads are used for computing prompts.
Modify these settings carefully, as this could lead to some extensive RAM or CPU usage if the values are set too high.