From time to time we get asked to provide an example for ChatGPT. Since this is just another web service, we can just handle it with the MBS Xojo CURL Plugin. But since we don’t like to block the user interface while ChatGPT processes the request, we use our CURLSMultiMBS class to run it in the background. Later when finished, it performs a delegate to call our Finished method to process the result.
The request is build with the JSONItem class to fill in the various values. We include a system message with the request, e.g. “Please translate text to English.” and then pass the text to translate in the user role. This way the user should not be able to provide instructions to the LLM in their text.
To test the project, please get an account for OpenAI. In the headers we need to pass your organization ID from and the bearer token you got there. Each query will use some tokens, but we limit the answer to 100 tokens.
Here is the full method to start the query:
Sub RunQuery()
Dim curl As New CURLSMBS
dim MessageSystem as new JSONItem
MessageSystem.Value("role") = "system"
MessageSystem.Value("content") = SystemField.Text
dim MessageUser as new JSONItem
MessageUser.Value("role") = "user"
MessageUser.Value("content") = UserField.Text
dim messages as new JSONItem
messages.Add MessageSystem
messages.Add MessageUser
dim j as new JSONItem
j.Value("model") = "gpt-4o-mini"
j.Value("max_tokens") = 100
j.Value("messages") = messages
curl.SetOptionHTTPHeader _
array(_
"OpenAI-Organization: " +OrganizationField.text, _
"Authorization: Bearer " +BearerTokenField.text, _
"Content-Type: application/json" )
curl.OptionPostFields = j.ToString
curl.OptionURL = "https://api.openai.com/v1/chat/completions"
call CURLSMultiMBS.SharedInstance.AddCURL curl, AddressOf Finished
End Sub
Here is the finished method, which gets called when the transfer is done. The script parameter is the CURL session reference number. We query debug messages and the result as JSON. We can then pick the output message from the JSON and show it in the field:
Sub Finished(curl as CURLSMBS, ErrorCode as Integer)
// we got a result asynchronously
ResultLabel.Text = ErrorCode.ToString+" "+curl.LastErrorText
DebugField.Text = curl.DebugMessages
ResultField.Text = curl.OutputData
RunButton.Enabled = true
Dim j As New JSONItem(curl.OutputData)
If j.HasKey("choices") Then
Dim choices As JSONItem = j.Value("choices")
Dim choice As JSONItem = choices.ChildAt(0)
Dim message As JSONItem = choice.Value("message")
Dim content As String = message.Value("content")
ResultField.Text = content
End If
Exception je As JSONException
// handle?
End Sub
Please try and see whether this can help you add ChatGPT to the projects. Translation is a good thing and you can even specify whether you like formal or informal text. We will include the example file with the next plugin for you to use.
You know I’m fine using plugins when I need to, but is there something CURLMBS offers here for interacting with the ChatGPT API that URLConnection does not?
The blog article is about showing how to do it with CURL functions.
While normal people may not see a difference, some clients may enjoy some of CURL’s options for proxy handling, TLS or DNS options and go around Apple’s NSURL security limitations.
I use chatgpt + customized AI models (GPT’s) focused on medicine within my app. Uses URLConnection inside a custom server control… works like a breeze, nothing else is needed.
I had trained some GPTs with my ChatGPT PLUS subscription for varios Bariatric Surgery Heathcare Centers. Using the OpenAI interface they are doing really well. It respect articule the instructions an mecer alucinates.
But as they are orientated for patients use, not all patients have a PLUS subscription, and with the free one I am having two limitations: 1) if the model used is gpt-4o the limited quantity of tokens available for Free subscription lets patients only be able to ask 5 to 8 questions máximum before arriving to the 4096 tokens máximum available, and 2) in the other hand, if the model used is GPT-3.5 the trained GPT not always adjusts itself to the instructions and some times alucinates.
So I want to solve this using OpenAI’s APIs
Just for test purposes I used the Xojo AI example. It works fine just using my APIkey with any of the Standard models (gpt-3.5-turbo or gpt-4o) but not with the trained model. The only one thing I changed was the model, from GPT-3.5-turbo to the ID of the personalized GPT.
I notice that you said that you have no problem on using custom trained GPTs.
Any advise in order to let me do it with my one GPTs ???
The ID of the model I understand is:
g-681649aaff14819186550380b4b66e6d
but I receive the following error:
The model g-681649aaff14819186550380b4b66e6d does not exist or you do not have access to it.
I also make different combinations on how to pass the ID Model, like:
j.Value(“model”) = “gpt-681649aaff14819186550380b4b66e6d”
j.Value(“model”) = “681649aaff14819186550380b4b66e6d”
j.Value(“model”) = “g-681649aaff14819186550380b4b66e6d-bari-center”
j.Value(“model”) = “gpt-681649aaff14819186550380b4b66e6d-bari-center”
And none of them works.
Anyone knows the way to avoid this in order to use a personal trained GPT model in a XOJO App ?
Thanks @MarkusR …
Please tell me if I understand the difference between a Custom GPT and a Playground Assistant. The former is to be used with the OpenAI interface and not with the APIs an the Assistants are similar ones but to be used using the APIs. And that may be that why the message I receive is that the Model (the custom GPT) is not accesible.
Am I right ?
i cant create a Custom GPT because i not have plus.
Playground is just the web page for testing all api interfaces, to see the request and response data.
i would expect that you then can select your model there.
oki, this is the issue : you cannot call custom GPTs (from chat.openai.com) by ID via API.
The custom GPTs you create inside ChatGPT (e.g. via the “Explore GPTs” tab) cannot be used directly via the OpenAI API, including in Xojo or any external application. These GPTs are currently only accessible within the ChatGPT interface itself, not via https://api.openai.com.
Use OpenAI Assistants API instead, to get the job done (recommended)
Use a detailed system prompt, each time
Lastly, you need to implement max-question per user per time interval as well, the cost can quickly accumulate other-wise.