Just for fun, I wanted to see how I could replicate a basic ChatGPT like experience using the tools available to me in the Power Platform. I will cover setting up the Power Automate flow to interface the OpenAI completions endpoint using GPT-4 Turbo and a Power App for the UI. In a future post, I will create another flow(s) to work with the Assistants API and take this to the next level.
A quick look at the app and how it works. The app will keep the chat context active using the completions endpoint until you want to clear it out. Like I said, it’s basic!
Here is what I used to get this up and running:
Power App
Power Automate flow with the HTTP connector
OpenAI account
OpenAI API key (credit card required)
Here is the layout of the flow and Power App
The key for this to work and keep the context of the conversation flowing is to pass all of the previous questions and responses back to the API each time a new question is asked. Yes, that can start to add up in terms of burning a lot of tokens for a large chat.
The input is converted to the following format and passed to the API using the app.
role: user
content: question asked
The response from the API is the same.
role: assistant
content: response from the API
Using the two, the collection is built and displayed in a gallery.
Power App setup –
ButtonQuestion OnSelect property does the following:
Set the button text
Set the varQuestion variable to the value of TextInputQuestion
Add the question and user role to the colResponses collection
Convert the colResponses collection to JSON
Cal the f Get HTTP flow passing in the collection
with the response from the flow, add the value(s) to the colResponses collection
Reset TextInputQuestion
Set the button text
GalleryRespnses properties:
Visible: If(IsBlank(ThisItem.content),false,true)
Text: ThisItem.content
X: If(ThisItem.role = “assistant”,40,5)
AutoHeight: true
Now for the flow! The flow is triggered from the app and passes a value.
Messages: the new question from the app and previous responses and questions (collection)
The Parse JSON action will take the Messages input and shape the JSON used in the Select Messages action.
The Select Messages action will use the output of the Parse JSON action to help form the correct input for the HTTP action.
Using a Compose action, the model and messages values are set. Note: in the example, I’m using GPT 4 Turbo preview.
The HTTP Request action uses the following values:
Method: Post
URL: https://api.openai.com/v1/chat/completions
Headers:
{
“Content-Type”: “application/json”,
“Authorization”: “Bearer @{outputs(‘Compose_API_Key’)}”
}
Body: output of the Compose Message Body
Another Parse JSON action is used to handle the response from the HTTP Request.
In the Select Response action, the output of Parse JSON Response is used to populate the role and content values.
The final action is to respond to the Power App. The respond to PowerApp step will not work if you use the dynamic content value to populate the response. You must use this expression: outputs(‘Select_Response’)
Again, if data is not sent back to the Power App, chances are you missed the sentence before this one.
As I mentioned at the beginning of this post, this is a basic example. If you see anything wrong with the process or places where it could be streamlined, please let me know.
