^for

A series of C# functions wrapping the ollama APIs, mainly for the UnityEngine

Prerequisite

Both the developer and the user's systems need to have a working ollama setup already:

1. Download and Install ollama
2. Pull a model of choice from the Library
   • Recommend gemma3:4b for general conversation

     ollama pull gemma3:4b

   • Recommend gemma3:1b for device with limited memory

     ollama pull gemma3:1b

   • Recommend deepseek-r1:7b if "reasoning" is needed

     ollama pull deepseek-r1:7b

   • Recommend nomic-embed-text for generating embeddings (vectors)

     ollama pull nomic-embed-text

In Unity, you need the Newtonsoft.Json package:

1. Unity Editor
2. Window
3. Package Manager
4. Add package by name
5. Name:

   com.unity.nuget.newtonsoft-json
   

6. Add

Then, simply download and install the .unitypackage from Releases

Features

The following functions are available under the Ollama static class; all functions are asynchronous

• List()
  • Return an array of Model, representing all locally available models
  • The Model class follows the official specs

Tip

You can use the families attribute to determine if a model is multimodal (see #2608)

Generate

• Generate()
  • The most basic function that returns a response when given a model and a prompt
  • You can also pass a Texture2D when using a model with vision capability
• GenerateStream()
  • The streaming variant that returns each word as soon as it's ready
  • Requires a callback to handle the chunks
• GenerateJson()
  • Return the response in the specified class / struct format

Important

You need to manually tell the model to use a JSON format in the prompt

Chat

• Chat()
  • Same as Generate(), but now 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
• InitChat()
  • Initialize / Reset the chat history
  • historyLimit: The number of messages to keep in memory
• SaveChatHistory()
  • Save the current chat history to the specified path
• LoadChatHistory()
  • Load the chat history from the specified path
  • Calls InitChat() automatically instead if the file does not exist

RAG

Retrieval Augmented Generation

• Ask()
  • Ask a question based on given context
  • Requires both InitRAG() and AppendData() to be called first
• InitRAG()
  • Initialize the database
  • Requires a model to generate embeddings
    • Can use a different model from the one used in Ask()

Important

Not all models can generate embeddings

• AppendData()
  • Add a context (eg. a document) to retrieve from

Demos

4 demo scenes showcasing various features are included:

• Generate Demo

  • List()
  • Generate()
  • GenerateJson()

• Chat Demo

  • InitChat()
  • ChatStream()

• RAG Demo

  • InitRAG()
  • AppendData()
  • Ask()

• Image Demo

  • Generate() with Texture2D