Skip to content

Latest commit

 

History

History
376 lines (277 loc) · 22.1 KB

README.md

File metadata and controls

376 lines (277 loc) · 22.1 KB
Raw

OpenAI WebUI

Latest version: 0.9.10 (20241217)

WebUI (streamlit-based) to ChatGPT and Dall-E's API (requires an OpenAI API key).

The tool's purpose is to enable a company to install a self-hosted version of a WebUI to access the capabilities of OpenAI's ChatGPT and DallE and share access to the tool's capabilities while consolidating billing through the OpenAI API key. Access to models is limited to those enabled with your API key.

Click on the links to see a screenshot of the GPT WebUI and the DallE WebUI.

Please see https://github.com/Infotrend-Inc/OpenAI_WebUI/blob/main/.env.example for details of possible values for the environment variables. Unless specified, even if a feature is not used, its environment variable should be set.

A pre-built container is available from our Docker account at https://hub.docker.com/r/infotrend/openai_webui

An Unraid-ready version is available directly from Unraid's `Community Applications``.

Note: this tool was initially developed in February 2023 and released to help end-users.

1. Description

The tool provides a WebUI to ChatGPT and Dall-E (that later one can be disabled).

The tool requires the use of an OpenAI API key to work. Check at https://platform.openai.com/account/api-keys to find yours.

Depending on your deployment solution (python virtualenv, docker image, or unraid), the deployment might differ slightly.

Once started, the WebUI will prompt the end user with a username. This username is here to make finding past conversations/images easier if you seek those; no authentication is associated with it.

ChatGPT (Text Generation) sidebar options (see "?" mark for specific details):

  • model: choose between the different ChatGPT models that are enabled.
  • role (user, system, assistant): define the role of the input text for tailored responses.
  • max tokens: controls the length of generated text with a maximum token setting (dependent on the model)
  • temperature: adjust the "surprisingness" of the generated text.

DALL-E (Image Generation) sidebar options (see "?" for specific details):

  • mode: "image" for the time being.
  • model: choose between the different DallE models that are enabled.
  • image Size: specify the dimensions of the images to be generated.
  • number of images (model dependent): number of images to generate
  • quality (model dependent): fine-tune image quality to meet your requirements.
  • style (model dependent): style of the generated images.

1.1. Supported models

We have added means to inform the end-user when a model is deprecated, legacy or current.

  • deprecated models are not available for use anymore.
  • legacy models will be deprecated at a specified date.
  • current models are available.

The tool will automatically discard known (per the release) deprecated models and inform the end user. Similarly, the tool will note when a model is legacy. Please update your model selection accordingly.

The models.json file contains the list of models supported by each release (as introduced in v0.9.3). The following table shows the models listed in this file as well as the release it was added to:

Mode Model Status Capability Notes From
DallE dalle-e-2 active 0.9.3
DallE dalle-e-3 active 0.9.3
GPT gpt-3.5-turbo active 0.9.3
GPT gpt-3.5-turbo-0125 active 0.9.3
GPT gpt-3.5-turbo-0613 deprecated Deprecated on June 13, 2024 0.9.3
GPT gpt-3.5-turbo-1106 active 0.9.3
GPT gpt-3.5-turbo-16k deprecated Deprecated on June 13, 2024 0.9.3
GPT gpt-3.5-turbo-16k-0613 deprecated Deprecated on June 13, 2024 0.9.3
GPT gpt-4 active 0.9.3
GPT gpt-4-0125-preview active 0.9.3
GPT gpt-4-0613 active 0.9.3
GPT gpt-4-1106-preview active 0.9.3
GPT gpt-4-32k deprecated 0.9.3
GPT gpt-4-32k-0613 deprecated 0.9.3
GPT gpt-4-turbo-preview active 0.9.3
GPT gpt-4-turbo active vision 0.9.5
GPT gpt-4-turbo-2024-04-09 active vision 0.9.5
GPT gpt-4o active vision 0.9.4
GPT gpt-4o-2024-05-13 active vision 0.9.4
GPT gpt-4o-2024-08-06 active vision 0.9.8
GPT gpt-4o-2024-11-20 active vision 0.9.9
GPT chatgpt-4o-latest active vision 0.9.9
GPT gpt-4o-mini active vision 0.9.7
GPT gpt-4o-mini-2024-07-18 active vision 0.9.7
GPT o1-preview active beta 0.9.8
GPT o1-preview-2024-09-12 active beta 0.9.9
GPT o1-mini active beta 0.9.8
GPT o1-mini-2024-09-12 active beta 0.9.9
GPT o1 active vision beta + untested
GPT o1-2024-12-17 active vision beta + untested

beta models provide limited control over features or capabilities. Those currently also do not support system role: if using prompt presets those might not work.

An untested model is added but is not able to be tested by the development team until we get access to the model (currently open to Tier 5 customers)

Once a model is deprecated, using it in your models list will have it discarded from the available list with a notification.

Similarly, if a used model is listed as legacy, a notification of the upcoming deprecation will be shown in the UI.

1.2. .env

The .env.example file contains the parameters needed to pass to the running tool:

  • OPENAI_API_KEY as obtained from https://platform.openai.com/account/api-keys
  • OAIWUI_SAVEDIR, the location to save content (make sure the directory exists)
  • OAIWUI_GPT_ONLY, to request only to show the GPT tab otherwise, shows both GPT and DallE (authorized value: True or False)
  • OAIWUI_GPT_MODELS is a comma-separated list of GPT model(s) your API key is authorized to use. See https://platform.openai.com/docs/api-reference/making-requests for more information.
  • OAIWUI_DALLE_MODELS is a comma-separated list of DallE model(s) your API key is authorized to use.
  • OAIWUI_USERNAME (optional) specifies a username and avoids being prompted at each re-run. The default mode is to run in multi-user settings so this is not enabled by default.
  • OAIWUI_GPT_VISION will, for compatible models, disable their vision capabilities if set to False
  • OAIWUI_IGNORE_EMPTY (required for Unraid) discard errors in case the following environment variables are used but not set.
  • OAIWUI_PROMPT_PRESETS_DIR sets the directory that contains prompt presets. If a directory is provided, it must contains at least one valid json file.
  • OAIWUI_PROMPT_PRESETS_ONLY sets the JSON file that contains valid settings to use for the OAIWUI_PROMPT_PRESETS_DIR presets.

Those values can be passed by making a .env file containing the expected values or using environment variables.

The .env file is not copied into the docker or unraid setup. Environment variables should be used in this case.

1.3. savedir

The OAIWUI_SAVEDIR variable specifies the location where persistent files will be created from run to run.

Its structure is: savedir/version/username/mode/UTCtime/<CONTENT>, with:

  • username being the self-specified user name prompted when starting the WebUI
  • version the tool's version, making it easier to debug
  • mode on of gpt or dalle
  • the UTCtime, a YYYYY-MM-DD T HH:MM:SS Z UTC-time of the request (the directory's content will be time ordered)
  • <CONTENT> is often a json file containing the details of the run for gpt, but also the different png images generated for dalle

We do not check the directories for size. It is left to the end user to clean up space if required.

1.4. password protecting the WebUI

To do this, create a .streamlit/secrets.toml file in the directory where the streamlit app is started (for the python virtualenv setup, this should be the directory where this README.md is present, while for other deployment methods, please see the corresponding setup section) and add a password = "SET_YOUR_PASSWORD_HERE" value to it.

When the WebUI starts, it will see of secrets.toml file and challenge users for the password set within.

1.5. Using "prompt presets" (GPT only)

Prompt presets enable the preparation of custom methods to answer "user" prompt by specifying some "system" and "assistant" settings. It is used by setting the OAIWUI_PROMPT_PRESETS_DIR to a folder containg .json files.

We have provided an example directory containing one pre-configured "prompt preset". The example directory is named prompt_presets.example and its content is the file shakespeare.json which guides the GPT answer in the English used by Shakespeare.

The structure of the used JSON file follows OpenAI messages' API structure and as such should be adhere to as closely as possible. It contains a series of messages that will be passed at the begining of new conversations to the GPT to set the role to system (the direction the GPT is expected to follow when answering) and/or the assistant (past conversations/expected knowledge) for that GPT conversation. The content section is expected to be a text type with the text to provide to the GPT.

For example, one of the prompt for the shakespeare.json example is as follows:

       {
            "role": "system",
            "content": [
                {
                        "type": "text",
                        "text": "You are a helpful assistant. You also like to speak in the words of Shakespeare. Incorporate that into your responses."
                }
            ],
            "oaiwui_skip": true
        }

The name of the prompt preset is directly related to the name of the file; if the file is title shakespeare.json, the prompt will be named shakespeare.

Creating new "prompt presets" should be a matter of duplicating the example and replacing the content within the file.

Another method consists of passing the prompt to the WebUI and setting the role accordingly, then running a query. The content saved within the savedir will contain a messages structure that matches the role and content sections shown above. Integrate that content within a new prompt presets JSON file.

Note that the oaiwui_skip is not passed to the GPT, but is used to remove the content from the chat history.

1.5.1. prompt presets settings

When using "prompt presets", it is possible to make the tool behave such that the end user can only use a single model with a set temperature and maximum requested tokens. This JSON settings file is used by pointing the OAIWUI_PROMPT_PRESETS_ONLY environment variable to the location of the file.

We have provided an example prompt_presets_settings-example.json file. This example file contains:

{
    "model": "gpt-4o-mini",
    "tokens": 3000,
    "temperature": 0.5
}

, which will:

  • use the gpt-4o-mini model (which must be in the OAIWUI_GPT_MODELS list of authorized models)
  • requests a maximum of 3K tokens for the GPT answer. The maximum value per model differs so the tool will error if the requested value is too high (note this is not the context tokens, which covers the entire chat)
  • set the temperature to 0.5. The temperature controls the randomness of responses, with lower values yielding more deterministic answers and higher values producing more creative and varied outputs (the range is 0 to 1)

2. Setup

2.1. Python virtualenv (poetry)

The virtualenv setup requires poetry and the setup is defined in the pyproject.toml file.

This mode is for use if you have python3 and poetry installed and want to test the tool.

  1. Create and activate your virtual environment (in the directory where this README.md is located):

    $ poetry install
$ poetry shell

  2. Copy the default .env.example file as .env, and manually edit the copy to add your OpenAI API key and the preferred save directory (which must exist before starting the program). You can also configure the GPT models you can access with ChatGPT and disable the UI for Dall-E if preferred. Do not distribute that file.

    $ cp .env.example .env
$ code .env

  3. Edit the code if desired, and when you are ready to test, start the WebUI.

    $ streamlit run ./OpenAI_WebUI.py --server.port=8501 --server.address=127.0.0.1 --logger.level=debug

  4. You can now open your browser to http://127.0.0.1:8501 to test the WebUI.

2.2. Docker/Podman

The container build is an excellent way to test in an isolated, easily redeployed environment.

This setup prefers the use of environment variable, using docker run ... -e VAR=val

  1. Build the container

    make build_main

  2. Run the built container, here specifying your OAIWUI_SAVEDIR to be /iti, which will be mounted from the current working directory's savedir and mounted to /iti within the container:

    docker run --rm -it -p 8501:8501 -v `pwd`/savedir:/iti -e OPENAI_API_KEY="Your_OpenAI_API_Key" -e OAIWUI_SAVEDIR=/iti -e OAIWUI_GPT_ONLY=False -e OAIWUI_GPT_MODELS="gpt-4o-mini,gpt-4,o1-mini" -e OAIWUI_DALLE_MODELS="dall-e-3" openai_webui:latest

If you want to use the "prompt presets" and its "prompt presets settings" environment variables, those can be added to the command line. For example to use the provided examples add the following to the command line (before the name of the container): -v `pwd`/prompt_presets.example:/prompt_presets -e OAIWUI_PROMPT_PRESETS_DIR=/prompt_presets and -v `pwd`/prompt_presets_settings-example.json:/prompt_presets.json -e OAIWUI_PROMPT_PRESETS_ONLY=/prompt_presets.json

If you want to use the password protection for the WebUI, create and populate the .streamlit/secrets.toml file before you start the container (see password protecting the webui) then add -v PATH_TO/secrets.toml:/app/.streamlit/secrets.toml:ro to your command line (adapting PATH_TO with the full path location of the secrets file)

With all the above options enabled, the command line would be:

docker run --rm -it -p 8501:8501 -v `pwd`/savedir:/iti -e OPENAI_API_KEY="Your_OpenAI_API_Key" -e OAIWUI_SAVEDIR=/iti -e OAIWUI_GPT_ONLY=False -e OAIWUI_GPT_MODELS="gpt-4o-mini,gpt-4" -e OAIWUI_DALLE_MODELS="dall-e-3" -v `pwd`/prompt_presets.example:/prompt_presets:ro -e OAIWUI_PROMPT_PRESETS_DIR=/prompt_presets -v `pwd`/prompt_presets_settings-example.json:/prompt_presets.json:ro -e OAIWUI_PROMPT_PRESETS_ONLY=/prompt_presets.json -v `pwd`/secrets.toml:/app/.streamlit/secrets.toml:ro openai_webui:latest

It is also possible to populate a .env file and mount it within the /app directory. Note that -v options still need to be applied for. For example, adapt the provided .env.docker.example file that uses /iti for its savedir and similar mount as the above command line for the "prompt presets" (but does not use the secrets.toml). The command line can be command line can be simplified as:

docker run --rm -it -p 8501:8501 -v `pwd`/.env.docker.example:/app/.env:ro -v `pwd`/savedir:/iti -v `pwd`/prompt_presets.example:/prompt_presets:ro -v `pwd`/prompt_presets_settings-example.json:/prompt_presets.json:ro openai_webui:latest

You can have the Makefile delete locally built containers:

$ make delete_main

2.3. Docker compose

To run the built or downloaded container using docker compose, decide on the directory where you want the compose.yaml to be, and place the following as the content of the file:

services:
  openai_webui:
    image: infotrend/openai_webui:latest
    container_name: openai_webui
    restart: unless-stopped
    volumes:
      - ./savedir:/iti
      # Warning: do not mount other content within /iti 
      # Uncomment the following and create a secrets.toml in the directory where this compose.yaml file is to password protect access to the application
      # - ./secrets.toml:/app/.streamlit/secrets.toml:ro
      # Mount your "prompt presets" directory to enable those are options
      # - ./prompt_presets.example:/prompt_presets:ro
      # Mount the "prompt presets" settings file to limit users to the model, tokens and temperature set in the file
      # - ./prompt_presets_settings-example.json:prompt_presets.json:ro
    ports:
      # host port:container port
      - 8501:8501
    environment:
      - OPENAI_API_KEY=${OPENAI_API_KEY}
      - OAIWUI_SAVEDIR=/iti
      # Adapt the following as best suits your deployment
      - OAIWUI_GPT_ONLY=False
      - OAIWUI_GPT_MODELS=gpt-4o
      - OAIWUI_GPT_VISION=True
      # Even if OAIWUI_GPT_ONLY is True, please set a model, it will be ignored
      - OAIWUI_DALLE_MODELS=dall-e-3
      # Uncomment and enter a value if you are using a single user deployment
      # - OAIWUI_USERNAME=user
      # Enable the user of "prompt presets" present in the mounted directory (must have a directory matching in the `volumes` section)
      # - OAIWUI_PROMPT_PRESETS_DIR=/prompt_presets
      # Enable the "prompt presets" setting (must have a file matching in the `volumes` section)
      # - OAIWUI_PROMPT_PRESETS_ONLY=/prompt_presets.json

In the directory where the compose.yaml is located, create a savedir directory (it will be mounted as /iti within the running container), and create a .env file that needs only contain the OPENAI_API_KEY=value entry. If using a secrets.toml file with a password=WEBUIPASSWORD content, uncomment the entry in the compose.yaml file.

As configured, the container will restart unless-stopped which also means that unless the container is stopped it will automatically restart after a host reboot.

Run using docker compose up -d

The WebUI will be accessible on port 8501 of your host.

2.4. Unraid

For Unraid users, a special build mode is available to get a container using unraid's preferred uid/gid, use make build_unraid to build it.

The pre-built container has been added to Unraid's Community Applications.

The configuration file contains many of the possible environment variables, as detailed in the .env section. Omitted from the configuration file:

  • a Path mapping a secrets.toml file to the /app/.streamlit/secrets.toml location within the running docker container (read-only recommended). Before setting this, create and populate a file with the expected value (as described in password protecting the WebUI). For example, if your appdata location for the OpenAI WebUI was /mnt/user/appdata/openai_webui in which you placed the needed secrets.toml file, the expected XML addition would look similar to: 
    <Config Name="/app/.streamlit/secrets.toml" Target="/app/.streamlit/secrets.toml" Default="/mnt/user/appdata/openai_webui/secrets.toml" Mode="ro" Description="WebUI password protection -- secrets.toml file must exist with a password variable" Type="Path" Display="always" Required="false" Mask="false">/mnt/user/appdata/openai_webui/secrets.toml</Config>

3. Misc

3.1. Notes

  • If you run into an error when starting the tool. Clear the streamlit cache (right side menu) or deleting cookies should solve this.

3.2. Version information/Changelog

  • v0.9.10 (20241217): Added o1 model (untested) following its API access availability
  • v0.9.9 (20241206): API changes to use o1-mini and o1-preview (tested)
  • v0.9.8 (20241010): Added o1-preview and o1-mini model (untested) + "prompt presets" functionalities
  • v0.9.7 (20240718): Added gpt-4o-mini and deprecated older 32k models
  • v0.9.6 (20240701): Added method to disable vision for capable models + added whole WebUI password protection using streamlit's secrets.toml method
  • v0.9.5 (20240611): Added support for vision in capable models + Added gpt-4-turbo models + Deprecated some models in advance of 20240613 + Updated openai python package to 1.33.0 + Decoupled UI code to allow support for different frontends.
  • v0.9.4 (20240513): Added support for gpt-4o, updated openai python package to 1.29.0
  • v0.9.3 (20240306): Simplifying integration of new models and handling/presentation of their status (active, legacy, deprecated) + Cleaner handling of max_tokens vs context window tokens + updated openai python package to 1.13.3
  • v0.9.2 (20241218): Keep prompt history for a given session + allow user to review/delete past prompts + updated openai python package: 1.8.0
  • v0.9.1 (20231120): Print streamlit errors in case of errors with environment variables + Addition of gpt-3.5-turbo-1106 in the list of supported models (added in openai python package 1.3.0) + added optional OAIWUI_USERNAME environment variable
  • v0.9.0 (20231108): Initial release -- incorporating modifications brought by the latest OpenAI Python package (tested against 1.2.0)
  • Oct 2023: Preparation for public release
  • Feb 2023: Initial version

3.3. Acknowledgments

This project includes contributions from Yan Ding and Muhammed Virk in March 2023.