ChatGPT UI

ChatGPT (or Claude / Gemini Pro) to host yourself, targeted at business-ey usecases that may be useful within a company, written in Elixir + LiveView

What is this?

The goal is to create a simple ChatGPT UI, for you and your company to self-host.

Features

Support for OpenAI GPT4, Google Gemini (through Vertex) and Anthropic Claude models (through bedrock)
"Scenarios" for repetitive tasks you find yourself doing a lot
Use your own OpenAI API key
Google OAuth authentication
Restrict access to certain email domains
Set which model to use
Code highlighting with highlightjs
Streaming messages support
Dark Mode
Automatic content truncating based on token count with bert-base-multilingual-uncased

Configuration

In config.exs, you can set the following keys:

config :chatgpt,
  title: "Elixir ChatGPT",
  enable_google_oauth: true,
  restrict_email_domains: false,
  allowed_email_domains: ["google.com"]

  default_model: :"gpt-4",
  models: [
    %{
      id: :"anthropic.claude-3-sonnet-20240229-v1:0",
      provider: :anthropic,  # <- provider can only be :anthropic, :google or :openai currently
      truncate_tokens: 100_000,
      name: "Claude 3 Sonnet"
    },
		%{
      id: :"gemini-1.0-pro",
      provider: :google,
      truncate_tokens: 100_000,
      name: "Gemini Pro (Google)"
    },
    %{
      id: :"gpt-4",
      provider: :openai,
      truncate_tokens: 8000,
      name: "GPT4"
    },
    %{
      id: :"gpt-3.5-turbo-16k",
      provider: :openai,
      truncate_tokens: 15000,
      name: "GPT3.5 Turbo 16k"
    },
    %{
      id: :"gpt-4-32k",
      provider: :openai,
      truncate_tokens: 30000,
      name: "GPT4 32k (EXPENSIVE!)"
    }
  ],

Then in runtime.exs:

# generate them at https://platform.openai.com/account/api-keys
config :ex_openai,
  api_key: System.get_env("OPENAI_API_KEY"),
  organization_key: System.get_env("OPENAI_ORGANIZATION_KEY"),
  http_options: [recv_timeout: 50_000]

# If you set `enable_google_oauth`, also set the following keys: (by default, uses the environment variables)
# check https://github.com/dwyl/elixir-auth-google/blob/main/create-google-app-guide.md on how to get these keys

config :elixir_auth_google,
  client_id: System.get_env("GOOGLE_AUTH_CLIENT_ID"),
  client_secret: System.get_env("GOOGLE_AUTH_CLIENT_SECRET")

# if you want to use AWS Bedrock for Anthropic models
config :chatgpt,
  access_key_id: System.get_env("AWS_ACCESS_KEY_ID"),
  secret_access_key: System.get_env("AWS_SECRET_ACCESS_KEY"),
  region: System.get_env("AWS_REGION")

To enable google vertex, set the GOOGLE_APPLICATION_CREDENTIALS_JSON env variable with the content of your service account credentials. You can also set GOOGLE_APPLICATION_CREDENTIALS to the filepath of your service account json file, or use service account based auth when using GKE/GAE.

export GOOGLE_APPLICATION_CREDENTIALS_JSON='{xxxxx}'

On Claude support

To enable Anthropic Claude models, you need to go to AWS Bedrock and request model access for the Claude models.

Then generate a new access key pair to use for authentication

On Gemini Pro support

To enable Google's Gemini models, you need to enable the Vertex APIs in the google cloud console.

If you need to generate a service account credentials json, go into IAM settings, create a new service account, then generate a new key. Important here is that you provide the service account with Vertex permissions.

Scenarios

This UI decided to go with "scenarios" instead of the usual "recent conversations" approach of the official ChatGPT.

"Scenarios" are preconfigured tasks that the AI should do, ready to be bookmarked and used whenever the task needs to be done. These can be things like:

Rewriting JIRA tickets
Grading user stories
Explaining code
Rephrasing based on corporate language

Currently, scenarios are set in scenario.ex and follow the %ChatgptWeb.Scenario struct:

      %{
        id: "generate-userstory",
        name: "📗 Generate Userstory",
        description:
          "Give me the content of a ticket, and I will try to write a user story for you!",
        force_model: :"anthropic.claude-3-sonnet-20240229-v1:0", # optional, if you want a specific scenario to always go through a specific model. Model has to be in `models` config
        messages: [
          %ChatgptWeb.Message{
            content:
              "You are an assistant that generates user stories for tickets. First, take the inputted text and give a summary if the entered text is a good userstory or not, with explanation why.\nThen, generate a proper user-story with the inputted text in the format of 'As a X, I want to Y, so that I can Z'.",
            sender: :system
          }
        ],
        keep_context: false
      }

messages are instructions to send to the LLM when the scenario is loaded, typically instructing the AI what it should do. Use :system messages to provide admin commands. The 3.5-turbo model is not paying so much attention to system messages, so experiment with using :user instead.
keep_context specifies whether the conversation context should be kept. If this is set to false (recommended), each message that is sent to the AI will be treated as a new conversation, the AI will have no context on previously sent messages.

Usage

Quick development setup:

First, install Elixir 1.14 or higher (1.13 or lower will not work). Then:

# Clone the project
git clone https://github.com/dvcrn/chatgpt-ui
cd chatgpt-ui

# Install Elixir
mix deps.get

# Install node dependencies
yarn install      # or npm install
mix assets.setup
mix assets.build

# Start the dev server
export OPENAI_API_KEY=<api-key>
export OPENAI_ORGANIZATION_KEY=<org-key>
mix phx.server

Usage / Deploying with Docker

First of all, generate a new random keybase: mix phx.gen.secret. A default is set, but you should really change this before deploying.

docker build . -t chatgpt
docker run -e SECRET_KEY_BASE=<keybase here> -e HOST="localhost" -p 4000:4000 chatgpt

The container (by default) expects the following env vars (unless you changed it in runtime.exs):

HOST
PORT (default 4000)
SECRET_KEY_BASE
OPENAI_API_KEY
OPENAI_ORGANIZATION_KEY

And if you use google oauth:

GOOGLE_AUTH_CLIENT_ID
GOOGLE_AUTH_CLIENT_SECRET

And if you use Anthropic through Bedrock

AWS_ACCESS_KEY_ID
AWS_REGION
AWS_SECRET_ACCESS_KEY

And if you use Google Vertex

GOOGLE_CLOUD_PROJECT_ID
GOOGLE_APPLICATION_CREDENTIALS_JSON

Status

Super work-in-progress