OpenAI integration in Nextcloud

⚠️ The smart pickers have been removed from this app as they are now included in the Assistant app.

This app implements:

• Text generation providers: Free prompt, Summarize, Headline, Context Write, Chat, and Reformulate (using any available large language model)
• A Translation provider (using any available language model)
• A SpeechToText provider (using Whisper)
• An image generation provider

⚠️ Context Write, Summarize, Headline and Reformulate have mainly been tested with OpenAI. They might work when connecting to other services, without any guarantee.

Instead of connecting to the OpenAI API for these, you can also connect to a self-hosted LocalAI instance or Ollama instance or to any service that implements an API similar to the OpenAI one, for example: IONOS AI Model Hub, Plusserver or MistralAI.
Make sure to use the OpenAI-compatible endpoint instead of the custom ones they provide.

⚠️ This app is mainly tested with OpenAI. We do not guarantee it works perfectly with other services that implement OpenAI-compatible APIs with slight differences.

Improve AI task pickup speed

To avoid task processing execution delay, setup at 4 background job workers in the main server (where Nextcloud is installed). The setup process is documented here: https://docs.nextcloud.com/server/latest/admin_manual/ai/overview.html#improve-ai-task-pickup-speed

Ethical AI Rating

Rating for Text generation using ChatGPT via the OpenAI API: 🔴

Negative:

• The software for training and inference of this model is proprietary, limiting running it locally or training by yourself
• The trained model is not freely available, so the model can not be run on-premises
• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model's performance and CO2 usage.

Rating for Translation using ChatGPT via the OpenAI API: 🔴

Negative:

• The software for training and inference of this model is proprietary, limiting running it locally or training by yourself
• The trained model is not freely available, so the model can not be run on-premises
• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model's performance and CO2 usage.

Rating for Image generation using DALL·E via the OpenAI API: 🔴

Negative:

• The software for training and inferencing of this model is proprietary, limiting running it locally or training by yourself
• The trained model is not freely available, so the model can not be ran on-premises
• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model’s performance and CO2 usage.

Rating for Speech-To-Text using Whisper via the OpenAI API: 🟡

Positive:

• The software for training and inferencing of this model is open source
• The trained model is freely available, and thus can run on-premise

Negative:

• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model’s performance and CO2 usage.

Rating for Text-To-Speech via the OpenAI API: 🔴

Negative:

• The software for training and inferencing of this model is proprietary, limiting running it locally or training by yourself
• The trained model is not freely available, so the model can not be ran on-premises
• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model’s performance and CO2 usage.

Rating for Text generation via LocalAI: 🟢

Positive:

• The software for training and inferencing of this model is open source
• The trained model is freely available, and thus can be ran on-premises
• The training data is freely available, making it possible to check or correct for bias or optimise the performance and CO2 usage.

Rating for Image generation using Stable Diffusion via LocalAI : 🟡

Positive:

• The software for training and inferencing of this model is open source
• The trained model is freely available, and thus can be ran on-premises

Negative:

• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model’s performance and CO2 usage.

Rating for Speech-To-Text using Whisper via LocalAI: 🟡

Positive:

• The software for training and inferencing of this model is open source
• The trained model is freely available, and thus can be ran on-premises

Negative:

• The training data is not freely available, limiting the ability of external parties to check and correct for bias or optimise the model’s performance and CO2 usage.

Learn more about the Nextcloud Ethical AI Rating in our blog.

🔧 Configuration

Admin settings

There is an "Artificial intelligence" section in the admin settings where you can:

• Choose whether you use OpenAI, a LocalAI instance or another remote service
• Set a global API key (or basic auth credentials) for the Nextcloud instance
• Configure default models and quota settings

Personal settings

There is an "Artificial intelligence" section in the personal settings where users can set their personal API key or basic auth credentials, as well as view their usage quota. Users can also choose to disable the Nextcloud Assistant even if the Assistant app is installed.