[F/OS] 🧠 Artificial Intelligence and OpenAI!

Gordon_Lu · January 15, 2023, 7:35am

Introduction

Do you want to chat with AI? Generate an image from your command? Well, this is the extension for you!

This is an extension that allows you to integrate OpenAI and GPT-3 with AppInventor. You can ask OpenAI questions, chat with it, and even generate images!

Stats

Built with: Extension Template

Released on: 2023-01-14T16:00:00Z

Version: 5

Documentation

Method blocks

Chat

Chats with OpenAI. The prompt parameter specifies what the query/question is, such as "What color is the sky?" or even better questions.

The prompt parameter tells the extension the question that you want to ask.

The model parameter defines the language model that you will use. Here are some possible combinations.

A1: gpt-3.5-turbo
A2: gpt-3.5-turbo-0301

Note: in order to use the models B1 to B4, you must sign up for a waitlist and wait for OpenAI for approval. Find the waitlist here: https://openai.com/waitlist/gpt-4-api

B1: gpt-4
B2: gpt-4-0314
B3: gpt-4-32k
B4: gpt-4-32k-0314

The apiKey parameter tells the extension the API key of your OpenAI account.

The maxTokens parameter limit the length of the answer that you will get. For example, in order to get an answer of around one or two sentences, set this as 250 to be safe. You may get a short or imcomplete answer if you set this too low. All free users can spend 150K tokens every minute.

Do you want an informative or a creative answer? The temperature parameter defines this. This should be a decimal between 0 and 1, with values closer to 1 returning more informative answers, and values closer to 0 returning more creative answers.

The role parameter specifies the role of the user - do you want to talk to the OpenAI bot as a user or an assistant?

Parameters: prompt = text, apiKey = text, maxTokens = text, temperature = number (double)

GenerateImage

Generates an image according to the prompt. The prompt tells the extension what the image is about, such as "Monkey holding a banana". The apiKey tells the extension your OpenAI API key. The size parameter defines the size of the returned image. Please only use helper blocks for this part.

Parameters: prompt = text, apiKey = text

ModerateText

The moderation API is a service provided by OpenAI that allows you to check whether a text contains inappropriate or offensive information. Put the text that you want to analyze in the input parameter, and supply the block with an API key.

Parameters: input = text, apiKey = text

Stream

Functions the same as the Chat block, but gets the response as chunks so that you do not have to wait until OpenAI finishes generating your response. This is like the "typing" view you see in ChatGPT. The id parameter is just for you to keep track of the response; you can set it to any integer. The response won't be affected by the ID parameter. For the other parameters please check the documentation for the Chat block.

Credit to @David_Ningthoujam for reference and his help.

Parameters: id = number (int), prompt = text, apiKey = text, maxTokens = text, temperature = number (double)

Event blocks

Error

This event is fired when an error has occurred.

Parameters: error = text, block = text

FinishedStream

This event is fired when all chunks of the stream are returned via the GotStream event.

GeneratedImage

This event is fired when the extension has generated an image! You can feed an Image component directly with this URL.

Parameters: imageUrl = text

GotStream

This event is fired when OpenAI has responded to your stream request. The response parameter is the responded chunk.

Parameters: id = number (int), response = text

ModeratedText

This event is fired when OpenAI has moderated the text you inputted in the ModerateText block! 'flagged' refers to whether the text contains indecent information in general.

categoryFlags includes a dictionary of 7 types of inappropriate information (hate, hate/threatening, self-harm, sexual, sexual/minors, violence, violence/graphic). These 7 are the keys of the dictionary. Corresponding to them are boolean values for whether the text has this type of inappropriate information. Sample response:

categoryScores includes the same 7 types as the keys, but corresponding to them are decimals from 0 to 1. Values closer to 0 mean that the text violates less in this category; vice versa.

Parameters: flagged = boolean, categoryFlags = dictionary, categoryScores = dictionary

RespondedToChat

This event is fired when the extension has responded to your question. The tokensSpent parameter indicates how many tokens you have spent from your OpenAI account in order to ask this question.

Parameters: response = text, tokensSpent = number (int)

Downloads

AIX: com.gordonlu.openai.aix (134.1 KB)

AIA: OpenAIV5.aia (146.6 KB)

📜 History

V4: com.gordonlu.openai.aix (124.7 KB)
V3: com.gordonlu.openai.aix (120.0 KB)
V2: com.gordonlu.openai.aix (114.2 KB)

FAQ

Q1: Why is my answer incomplete? Is this a bug in the extension?

A1: No, this is not a bug. Try to increase your maxTokens parameter for this. The maxTokens parameter judges how long the response should be, and with lower tokens, you get shorter, or maybe incomplete, answers.

Q2: How do I obtain an API key?

A2: @Dayron_Miranda was generous enough to give us a few steps. Not all countries support OpenAI though, for example, we Chinese do not get to use OpenAI.

Go to the OpenAI website (https://openai.com/).

Click on the “Developers” button in the top right corner of the page.

Click on the “API” button in the top right corner of the page.

Click on the “Get API Key” button in the top right corner of the page.

Fill in the required information and click on the “Create API Key” button.

Once you have your API key, you can test GPT chat by using one of the many libraries or sample codes available on the OpenAI website, or by making a call to the API endpoint directly with the API key. In this instance, copy the API key for this extension to work.

Q3: Is there a quota for this?

A3: Yes, unfortunately. Please see Pricing for details.

Q4: Why do I get a "No value for choices" error?

A4: This is caused maybe because you include a special character. Do not use double quotes in your prompts ("), use single quotes ('). Also check if you completed the setup correctly.

Q5: Where is the source code?

A5: It's in this repository in GitHub.

Credits

Thanks again to @Kumaraswamy for guidance and help in creating the extension, also @shreyash for his tutorials in creating the extension! (it's here: Learn how to create extensions - Resources - Kodular Community)

Also, special thanks to @Dayron_Miranda, @AyProductions, @Darko and a donor who wished to remain anyonymous, who all provided an API key for me to test this extension, and also participating in the beta-test program himself. Shoutout to him! Without his help, this extension would not be possible. (OpenAI is not supported in Hong Kong)

Please, if you like this extension or if you want to support me, click the beautiful Like button of this extension, this took two whole mornings and one whole afternoon to make, and I genuinely appreciate your kindness. Have a nice day

Abbas_Gasri · January 15, 2023, 1:43pm

Thank you What about Unsupported countries can you give me ApiKey ?

Gordon_Lu · January 15, 2023, 10:25pm

Unfortunately, no. API keys are not meant to be shared publicly with a lot of people, and me myself is in an unsupported country. I am relying on someone else's key

H_Peter_Alesso · January 15, 2023, 10:36pm

Gordon,
Great Job!
Can you provide the source code for the image generator?
Peter

Gordon_Lu · January 15, 2023, 11:02pm

Thank you!

Check the link in GitHub. The full extension source code is there. Check lines 98-145.

manyone · January 16, 2023, 8:55am

you may have misled us a bit.

i started it up after providing my API key but i couldn't get it to return any good value (see my first 4 entries below).

so i started looking at the java code and it showed the set up was for connecting to GPT-3 - not ChatGPT! i realized that my trial subscription to GPT-3 had expired after my trial in the summer so i established a paid account and as soon as i did, the app started to work again - the AI was starting to respond. However it was obvious i wasnt connected to ChatGPT from its response to my prompt "make me a poem telling me why i should learn python" - sure it responded with a paragraph - not a poem - and some reasons were good and it seemed to have problems with "\n" (newline)! and it charged me 103 tokens!

look at how the real ChatGPT responded:

and it's still free! (use this link for chatgpt: https://chat.openai.com/chat)

this isnt really a connection to ChatGPT but only to GPT-3. also according to the java code, it was using Ada, the cheapest language model ($0.0004 per 1k tokens generated). most everyone else uses Davinci the most powerful model ($0.0200 per 1k tokens generated). as far as i know OpenAI hasn't released the API yet for ChatGPT. And it's still offering a free trial.

just my observation.
(this article explains the difference between GPT-3 and ChatGPT:
OpenAI API Pricing in Words per Dollar | by Ivan Campos | Sopmac Labs | Jan, 2023 | Medium)

Gordon_Lu · January 16, 2023, 9:05am

Hi manyone, thank you for your comments. I will try to explain everything, hopefully this can shed some light.

The typo is because you set your tokens too low; the API may cram the answer within the token limit and result in an incomplete answer.

Make sure you disable informative.

Hmm...I'm checking, but my beta tester said that he is using a free trial, in fact, for $18. I changed the model to Ada because other users told me that Davinci is too expensive - one call costs $0.02. Calling Davinci for 900 times may end your free trial. However, I will consider adding different models for the extension.

I may have mistaken ChatGPT and GPT-3. I'll investigate into this. Thanks.

I couldn't decide on this, since OpenAI decides how many tokens you would have to spend. However, a factor of this is the length of the answer. If you get a longer answer, you would have to spend more tokens.

fahadboss10 · January 16, 2023, 10:41am

I'm getting "No value for choices".
Even on your provided AIA, the same error is coming.

Gordon_Lu · January 16, 2023, 10:47am

Hi, did you provide a valid API key? Make sure you added a valid API key from OpenAI. Share with me your project in PM if you still find the error. Check the FAQ Q4 for this as well.

fahadboss10 · January 16, 2023, 10:50am

Yes, I provided the API.

Gordon_Lu · January 16, 2023, 10:59am

I'm contacting you in PM. Hope we can sort out the errors.

EDIT: Sorted out, this user has their account expired before today. Please check and make sure your accounts are still available.

Gordon_Lu · January 16, 2023, 12:08pm

Hi @manyone,

@fahadboss10 has confirmed that the extension is working. Can you try again and see if there are errors?

manyone · January 16, 2023, 3:26pm

There's nothing wrong with the extension. It works. I'm just saying that it's misnamed. The title implies that it's connecting to ChatGPT but it is actually connecting to GPT-3.
So you cannot really do all those wonderful things that you read about that you can do with ChatGPT.

Abbas_Gasri · January 17, 2023, 4:43pm

No value for choices

Gordon_Lu · January 17, 2023, 10:36pm

Is your API key expired already?

Mobile.arsmiha4_Dori · January 19, 2023, 7:31am

How to change the model from ada to davinci?

Gordon_Lu · January 19, 2023, 7:51am

It's not possible with this version yet, but I will add this feature in the next release.

Corresponding to @fahadboss10's question about how to format the answers with new lines, it's still not possible with this release, as the result returned from the OpenAI server adds line breakers () in weird locations. However, I'll try to figure out a way to sort things out.

Gordon_Lu · January 19, 2023, 11:38am

Hello everyone, I hereby present you the version 2 update of this extension.

New features:

Better error detection. The "no value for choices" error should be eliminated, and should be related to an error message including more detail, such as, "You exceeded your current quota, please check your plan and billing details.".
Choose your models. Previously I have set the only model to Ada, but now you can choose three more models, i.e. Babbage, Curie and Davinci.
Renamed extension to OpenAI to stop the confusion around. @Taifun, please kindly update your directory and @Peter, please update the links in your replies in other communities of this extension. Please remember to update the old extension, convert the blocks to the new one and delete the old extension. Thank you for your kindness.

Plans for next release:

Include options to resize images to 256 square and 512 square.

@manyone Is this OK now?

manyone · January 20, 2023, 2:51am

yes - it's looking good thank you - it's nice to be able to select the language model.
i want to suggest something too - the informative parameter seems too restricitive if there are only 2 values - straight and exact answers or "mildly creative" answers. i looked at the code and apparently, this is passed to gpt-3 as the temperature parameter which has a value, respectively, of 0 for exact answers and 0.9 for creative answers. i was wondering if it makes sense to replace the informative parameter with the temperature parameter - so the user can select the wildness of the responses as he wishes. it can be presented as a slider , 0.0 to 1.0 (extremely wild to boring).
also, the first demo didn't have an example for image generation. will the new demo show it this time?

Gordon_Lu · January 20, 2023, 3:02am

It's on Screen2; please check again.