Nea (macOS)

100% SwiftUI GPT/LLM Client with Prompt Engineering support. With Microsoft's Azure OpenAI functionality.

General Query View	Prompt Studio

Development Roadmap

Follow updates here

Table of Contents

• Requirements
• Swift Packages & Credits
• Setting Up Azure OpenAI
• Persisting Chat Messages
• Guide WIP
• FAQ
• More Previews
• TODO

Requirements

• macOS 12.4+ Build passing 🟢

Swift Packages & Credits

• Granite
  • This app uses my own design pattern. A decoupling guide will be provided in the repo above. But, should be pretty straightforward overall.
• VaultKit
• SwiftGPT by @SwiftedMind
• GPT3-Tokenizer by @aespinilla
• swiftui-markdown by @jaywcjlove
• Ink by @JohnSundell
• LaunchAtLogin by @sindresorhus
• PureSwiftUI by @Brett-Best

Locally Diverged

• [SandKit] is a private repo at the moment, diverged a localized version for use. It houses SwiftGPT and swiftui-markdown from above as well. This kit handles custom prompt/command routing for the applicatoin.

Removed

• [MarqueKit] is a private repo that handles encryption this has been removed and comments show where encryption can take place to safely store data like API Keys.

Setting Up Azure OpenAI

1. Enable Azure in Settings

2. In SandGPT, setup the config to match your Azure OpenAI Resource/Deployment/ApiKey

static var AZURE_CONFIG: ChatGPTAzure.Config {
    .init(apiKey: "...",
          apiVersion: "2023-05-15", //apiVersion not necessary for initialization
          resourceName: "...",
          deploymentName: "...")
}

Persisting Chat Messages

Repo will be updated to include this out of the box, front-end side, in the future.

Back in SandGPT:

Nea/Services/Sand/Models/SandGPT.swift

Line 48 in 24fbcfb

func ask<E: EventExecutable>(_ prompt: String,

Simply build a flow to store the type ChatMessage prior to using the Swift Package.

private var messages: [ChatMessage] = []
    
func ask<E: EventExecutable>(_ prompt: String,
                             withSystemPrompt systemPrompt: String? = nil,
                             withConfig config: PromptConfig,
                             stream: Bool = false,
                             logMessages: Bool = false,
                             event: E) {

Append the user prompt with the role user.

self.messages.append(.init(role: .user, content: prompt)) //add user prompt

Append the chat-bot's response with the role assistant.

reply = try await client.ask(messages: self?.messages ?? [],
                                                         withConfig: config)
                            
DispatchQueue.main.async { [weak self] in
    self?.isResponding = false
}
    
if logMessages {
    self?.messages.append(.init(role: .assistant, content: reply))
}
    
event.send(Response(data: reply, isComplete: true, isStream: false))

The role .system is good for setting a persona prior to message collection.

self?.messages.append(.init(role: .system, content: "Act as..."))

Linking the reset() function in SandGPT() to clear messages is helpful too:

func reset(messages: Bool = false) {
    self.currentTask?.cancel()
    self.reqDebounceTimer?.invalidate()
    self.replyCompletedTimer?.invalidate()
    self.isResponding = false
    
    if messages {
        self.messages = []
    }
}

Guide

PaneKit

• Window resizing and size management occurs here.

Declaritevly update a single window's size whenever an action requires

state.pane?.display {
    WindowComponent(WindowComponent.Kind.query)
    
    if addResponse {
        WindowComponent(WindowComponent.Kind.divider)
        
        WindowComponent(WindowComponent.Kind.response)
        
        WindowComponent(WindowComponent.Kind.spacer)
        
        WindowComponent(WindowComponent.Kind.shortcutbar)
    }
}

InteractionManager

• Popups and Hotkey observation/registration

Example of Using PopupableView to easily trigger popups in any window instance.

PopupableView(.promptStudio,
              size: .init(200, 200),
              edge: .maxX, {
    RoundedRectangle(cornerRadius: 6)
        .frame(width: 60, height: 60)
        .foregroundColor(Color(hex: promptColor))
}) {
    ColorPicker(hex: $promptColor)
}
.frame(width: 60, height: 60)

FAQ

Why chat completions over completions for prompts?

• No specific reason besides, finding it to provide better results for my own needs.
• Completions endpoint capability will be added soon along with a toggle to switch between.

More Previews

Advanced Tuning	Helper Tab

Using Commands	Switching Commands

TODO

• Clean up Azure OpenAI integration in SwiftGPT.
• Integrate Azure and chat history properly into main App. Cleanup README after.
• Completions support along with chat completions.