ovos play utilizes the LLMs (Large Languag Models) created by OpenAI for its AI supported features and you can choose between the regular OpenAI or OpenAI on Azure which allows using servers in Europe. (for further details on data privacy / GDPR please see https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy).
The region in which the service is hostet does not affect if the data is used for training the LLMs. According to OpenAI, no sent data is used for training the LLMs when using the API provided.
Configuring OpenAI
1. Set up an OpenAI API Account
In order to configure OpenAI you need to subscribe to the OpenAI API. A ChatGPT Subscription cannot be used for ovos play. To setup an account for OpenAI API go to https://platform.openai.com/.
2. Create an API Key
Once you created your account creation you can create API keys on https://platform.openai.com/api-keys. To use the ovos play feature, we need you to provide an API key.
3. Configure a billing method
Using the OpenAI API requires you to configure a billing method for your account. We recommend using pay-as-you-go billing with a monthly limit of 350$. You can adjust payment warnings and limits under https://platform.openai.com/settings/organization/limits.
The cost of an OpenAI request varies depending on the type and context length. Approximate guidelines are:
Creation of an image: ~$0.08
Creation of a library structure: ~$0.04
Creation of maps: ~$0.02
Voice-over creation: ~$0.0015 for 100 characters
Simple search query: ~$0.0002
4. Add at least 5$ credit balance
The OpenAI API accounts work on tiers, which are leveled up for your account depending on the payment. To use the models that ovos play requires, you need to add at least 5$ to your credit balance in order to get started.
5. Send the API Key
Once everything is configured, please send your API key to your customer success manager in a secure way (e.g. via a password manager). The AI usage is then configured in ovos play and we will let you know once you can start using the AI features.
Configuring Azure OpenAI
Please note that OpenAI on Azure does currently not support Dall-E in the EU region. We do not yet support generation of images with Dall-E in ovos play when using Azure OpenAI.
It is possible to use Azure OpenAI Service with the ovos play API features. You can then (almost) freely choose which server location you want to utilize for the ovos play AI features. Not all regions support all models. Support is therefor limited.
1. Setup an Azure OpenAI Services instance
The first thing you have to do is to setup an Azure OpenAI Services instance. You can do this by going to the Azure OpenAI Services page. (or https://portal.azure.com/#view/Microsoft_Azure_ProjectOxford/CognitiveServicesHub/~/OpenAI)
During this step, you'll have to select the region to use. Currently, the most capable region in EU is Sweden.
Not all regions support all models. Support is therefor limited. You can look at https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models#model-summary-table-and-region-availability for more information. However, the information on this page is also not always up to date.
2. Create model deployments
For each model which should be used, you have to setup a separate deployment.
This is done in the Azure OpenAI Studio for your instance.
Even though you may see all available models - these are not necessarily supported in the region you have chosen. You'll receive an Error message from Azure when trying to save the deployment if the model is not supported in the region.
You need to configure the following models (with these names):
'gpt-4o'
'gpt-4o-mini'
'tts-1' (used to generate voice overs for contents)
'whisper' (used for audio transcription when using the AI task template)
'dall-e-3' (used for image generation)
3. Configure each model for ovos play
Please provide the following information to your customer success manager:
The instance name for each model (you may deploy models on different instances)
The deployment name for each model
The API key for each deployment/instance (you may use the same key for the same instance)
(optional) Custom endpoint for each model