ChatGPT attempts new questions from the questions channel

[tmcdonnell2]

Code execution from user perspective/birds-eye view:

ROSY PATH

Step 1) Enter your question into the questions forum. The question is formed from the title and question message. The title is always include and the message takes up the remaining space we have set for messages.
Additionally, the tags selected by the user are appended to the front of the question to provide more context to ChatGPT. From just a quick look at questions in the forum on the main discord, people ask questions without proper context, much like how titles doesn't include anything about Java since its implied by the asking in this discord. However, users generally seem good at adding the tags for the question.

Step 2) If a suitable question is found, it is sent to ChatGPT. Once the response returns it is checked to see if it was a bad answer. The answer is sent as an embed of an explanation message: it explains that the next message is AI generated, someone should be long soon, and you can use /chatgpt to ask the AI more questions. All this is sent AFTER the original help thread messages (format your code like this, provide details, close the thread when done).

ERROR PATH

Should at any point an error be produced, then all is returned is a message about how you have use /chatgpt to ask questions while you wait for a human. It is like the normal path except for an answer from ChatGPT.

REASONS FOR ERRORS

• Timeout (most common)
• ChatGPT unable to respond (a basic check for "As an AI language model" since this usually means a non-response)
• Combined the question title and question message are both combined < 20 characters since it is unlikely that this will be enough context for ChatGPT to work with and it either won't respond or ask for more context.

[Zabuzard]

Mind sharing some screenshots of the flow? Much easier to look at than starring at the code only.

Also, you gotta resolve that merge conflict :) Seems you worked from the wrong branch and now you have the other unrelated changes in this branch as well, which causes the merge conflict. You have to discard those changes and correctly rebase on develop.

[Zabuzard]

mind sharing up2date pictures of the conversation and the error cases? ideally put them into the description of the PR.

not sure if u did that, but i wanted to have no message at all (just silence) if chatgpt has an issue - did u do that? or does it still respond with some fallback "there was an error blablabla"

[tmcdonnell2]

mind sharing up2date pictures of the conversation and the error cases? ideally put them into the description of the PR.

not sure if u did that, but i wanted to have no message at all (just silence) if chatgpt has an issue - did u do that? or does it still respond with some fallback "there was an error blablabla"

I updated the description to show these paths. This shows what messages are sent and you can see if you like it or not.

[Zabuzard]

never use exceptions for logic that u could also check with simple if-else. in particular here, the Optional API has a very rich orElse API, make use of it.

answer.orElseGet(() -> {
  logger.warn(...);
  return useChatGptFallbackMessage(...);
});

[tmcdonnell2]

While I appreciate what the Optional API can do, I think how I have this setup will require some reworking to fix. I use useChatGptFallBackMessage to create a RestAction<Message>. This is required since inside that function it used mentionGuildSlashCommand which creates a RestAction<String>. From looking at the RestAction documentation, this is no way to get the content of the message out (beyond reflection or blocking the thread which is not desired from other code I wrote). If I were able to get it out as a String then the other method would have to be rewritten. For now, I have replaced it with a simple check if the optiona is empty and then logging and returning the fallback message. Let me know if there is some other way I should be doing this.

[Zabuzard]

i suppose questions are rarely max length, are they? try to go for something that is more common in practice, maybe max / 2?

[tmcdonnell2]

You would be surprised just how fast 200 characters gets filled up. Just taking some sample questions that are on the forum right now they all are around 200 or exceed 200 characters (MAX_LENGTH as defined of now). If the person includes almost any code (and with Java being verbose) the characters are destroyed. It might even be good if we carve out code in the future and just try to get to the human language bits of the question.

The other bit is that I add the tags to the question if a question is too short thus pushing up question lengths.

[Zabuzard]

i think it would be easier to read if the order would be changed here. like:

questionBuilder.append(title).append(" ");
String shortQuestion = question.substring(0, questionBuilder.length());
questionBuilder.append(shortQuestion);

Then the comment is also not needed anymore, since its understandable.

U have to see how that if fits into the mix though.

[tmcdonnell2]

questionBuilder.append(questionTitle).append(" ");

if (originalQuestion.length() > MAX_QUESTION_LENGTH - questionBuilder.length()) {
    originalQuestion = originalQuestion.substring(0, MAX_QUESTION_LENGTH - questionBuilder.length());
}

questionBuilder.append(originalQuestion);

This is what I got after some refactoring. More or less readable?

[Zabuzard]

in the past, this helper method returned the explanation message, now it returns the chat gpt message.
please double check all its current users to make sure thats okay with them

[tmcdonnell2]

It still returns the explanation message. It only has one usage which is in HelpThreadCreatedListener:

private void handleHelpThreadCreated(ThreadChannel threadChannel) {
        helper.writeHelpThreadToDatabase(threadChannel.getOwnerIdLong(), threadChannel);

        Runnable createMessages = () -> {
            try {
                createMessages(threadChannel).queue();
            } catch (Exception e) {
                logger.error(
                        "Unknown error while creating messages after help-thread ({}) creation",
                        threadChannel.getId(), e);
            }
        };

        // The creation is delayed, because otherwise it could be too fast and be executed
        // after Discord created the thread, but before Discord send OPs initial message.
        // Sending messages at that moment is not allowed.
        SERVICE.schedule(createMessages, 5, TimeUnit.SECONDS);
}

This is called by another function onChannelCreate which checks to ensure that what is being created is a help thread.

Is this what you mean? I think this works since we want to send this along with each new helper thread created. The messages created also appears to be the ping to the helpers and the explanation messages. Seems like a good place to extend the functionality. But I can appreciate that it might need a new place to live. But it seems like it would still be part of the queue() mentioned in the above code. Let me know what you think.

[sonarqubecloud]

Kudos, SonarCloud Quality Gate passed!   

0 Bugs
0 Vulnerabilities
0 Security Hotspots
0 Code Smells

0.0% Coverage
0.0% Duplication

[Tais993]

What if the response is over the character limit? ChatGPT often gives really big responses, so I can't imagine this to not happen.

[marko-radosavljevic]

Yup, that's already an issue with /chatgpt that is on master. It silently crashes if it's over the character limit.

And just cutting the answer to fit the limit is a bad solution, obviously. What's the point of getting 80% of the answer, if it's a code example or a step-by-step guide? Code won't work, guide is incomplete and useless.

So one solution is to prompt it to fit the answer within limits, and test it to make sure it works well. With a hard cutoff if a limit is reached.

Another would be to post lengthier answers as a file.

[Tais993]

What if we simply ask ChatGPT to limit the answer to a certain limit. (within 4000 characters)

ChatGPT can still exceed that limit, so we'd still require a fallback.
But it should work for most cases, and we don't cut the answer of or anything.

[marko-radosavljevic]

Yeah, that's one of my proposed solutions.

We either tell it to fit the answer withing the limit (limit - 500 for optimal results), with a fallback. Or we let it post the full answer as a file.

I like both of them. First one is short and neat, no files. My only worry is that the quality of the answers might lack. For example, can you fit a coherent guide how to install and setup something, with examples, within the limit.

[marko-radosavljevic]

Tested few examples, it seems embed length is enough for most guides with examples.

My worry with embeds is that they are too narrow for code, it might ruin the formatting. Since code would be needed in almost every answer, it's something to consider.

And 2000 char limit or regular messages is likely not enough for bigger answers. Close, but not quite there.

[tmcdonnell2]

Tested few examples, it seems embed length is enough for most guides with examples.
My worry with embeds is that they are too narrow for code, it might ruin the formatting. Since code would be needed in almost every answer, it's something to consider.

And 2000 char limit or regular messages is likely not enough for bigger answers. Close, but not quite there.

Thus, we should fit answers in a file?

[marko-radosavljevic]

Try embeds first, for cosistency. Post a picture of a more complicated answer with code examples.

Files are the last restort, since they are kinda ugly and weird UI wise. You know, getting a file in response to a bot command.

[Tais993]

A file is fine for PC, you have good UX then, not as good as just a message but it's fine.

On phone, yeah uh, have fun opening notepad on your phone to read it.

[github-actions]

This pull request is stale because it has been open 30 days with no activity. Remove stale label, comment or add the valid label or this will be closed in 5 days.