Ollama for Unity

A series of functions wrapping the ollama APIs, mainly for use in Unity

Prerequisite

The user's system needs to have a working ollama setup already:

1. Download and Install ollama
2. Pull a model of choice from the model library
   • Recommend llama3 for general text conversation

     ollama pull llama3

   • For coding-related inquiries, also try out stable-code

     ollama pull stable-code

   • Recommend llava for image captioning

     ollama pull llava:13b

   • Recommend mxbai-embed-large for embeddings

     ollama pull mxbai-embed-large

In Unity, you also need the Newtonsoft.Json package:

• Unity Editor -> Window -> Package Manager -> + -> Add package by name -> com.unity.nuget.newtonsoft-json

Features

All functions below are asynchronous. Simply call the functions under the Ollama class. Two demo scripts showcasing how to use each function are included.

• List()
  • Return an array of all available models that you can pass to other functions below
• ListCategorized()
  • Return two arrays, regular text models (eg. llama3) and multimodal models that can handle image inputs (eg. llava), separately instead
• Generate()
  • The most basic function that returns the response when given a prompt
• GenerateStream()
  • The stream variant that passes each word as soon as it's ready (like ChatGPT). Requires a callback to handle the texts.
• GenerateWithImage()
  • Generate a response based on an image. Requires a multimodal model (eg. llava). Supports UnityEngine's Texture2D directly.
• GenerateWithImageStream()
  • Same as above
• GenerateJson()
  • Give a class / struct format for the model to reply in
  • Important: You need to manually tell the model to reply in JSON format in the prompt as well
• Chat()
  • Same as Generate(), but with the memory of prior chat history, thus allowing you to further ask about previous conversations. Requires either InitChat() or LoadChatHistory() to be called first.
  • Example:

    >> Tell me a joke
    "..."
    >> Explain the joke
    "..."
    

• ChatStream()
  • Same as above
• ChatWithImage()
  • Same as above
• ChatWithImageStream()
  • Same as above
• InitChat()
  • Initialize / Reset the chat history
  • historyLimit: The number of messages to keep in memory (defaults to 16)
• SaveChatHistory()
  • Save the current chat history to the specified path (defaults to Application.persistentDataPath)
• LoadChatHistory()
  • Load the chat history from the specified path (defaults to Application.persistentDataPath). Will simply initialize if the file does not exist.

Note: All model-related functions (eg. create, copy, pull, etc.) will not be implemented

Retrieval Augmented Generation

^experimental

Based on ChromaDB

• Ask()
  • Ask a question based on the given context. Requires the following to be called first.
• InitRAG()
  • Start the ChromaDB server and initialize the database
  • pythonPath: The path to the Python folder. See Prerequisite.
  • authToken: ChromaDB allows you to set a password to prevent other users from accessing the database.
• AppendData()
  • Give the context (eg. a document) to perform RAG on

Prerequisite

This repo comes with an Editor script that helps you install the necessary self-contained Python environment for running the ChromaDB.

Simply go to the Editor, click Ollama in the toolbar, then click Obtain Python. Specify a path to store the environment, then click the 2 install buttons.

Afterwards, when you call InitRAG, remember to pass in the path to the python folder.

W.I.P

• AskChat()
  • Ask a question based on the given context as well as chat history