oak wedge
I am developing a Twitch Chatbot that integrates the GPT-3 API with the Twitch API.
Our goal is to eventually incorporate the chat gpt API once it becomes publicly available.
We would appreciate it if you could try out the chatbot and see if you can find any issues or areas for improvement.
Here is a link to the chatbot on Twitch: https://www.twitch.tv/all_in_89
Thanks!
GPT-3 API Twitch Chatbot
scenic night
I would like to use this for a YouTube video.. how would I implement this?
oak wedge
this is specifically coded for Twitch but I am sure there is a way to code it to respond to youtube comments. You would have to use google's API
north ivy
kinda fun 😉 if the answer is too long, twich doesn't display all the text
oak wedge
@north ivy I fixed this!
north ivy
<@87125967042011136> I fixed this!
What happend when several people speak at the same time?
oak wedge
What happend when several people speak at the same time?
It ques them up and answers them in order
north ivy
Do you think the answers of the IA will be related with previous answers provided to others people?
oak wedge
no, the davinci text model that the api currently offers does not remember prevoius responses
I hope to incorporate chatgpt into the bot when they realease and api for it, and when they do i am sure there will be documentation as to how to work the memory
north ivy
Okey ^^ i find it interesting because i want to do a discord bot where people can chat with the bot
Kinda the same concept you made but for discord
oak wedge
im sure that can be done
oak wedge
bump
rare crag
no, the davinci text model that the api currently offers does not remember prevo...
You could potentially add a realtime buffer upto a certain number of tokens to include in the prompt as part of a "recent" conversation. It may require some prompt engineering, but I can't see it being a problem other than it could become expensive to maintain in a busy chat.
oak wedge
thats an interesting concept. what would be the best syntax to pass a "recent conversation" to davinci?
@rare crag
rare crag
thats an interesting concept. what would be the best syntax to pass a "recent co...
At a sort of high level, since I've no code to draw from:
New_Prompt = {Some_new_Prompt}
Bot.Ask(Some_Dict + ". The text above is from the previous conversation, please refer to it when responding to the next sentence below. " + New_Prompt)```gotcha thanks
rare crag
Might wanna use the Twitch API to check for "reply" messages IN those previous 5 messages, as they may not be in a 5 message range, but they may be more useful to the conversation.
oak wedge
I just created a dictionary and concatenated the query and the OpenAI API response into the value of the dictionary. then I limited the dictionary to 1000 characters
has_before = previous_response.get(user)
if has_before is None:
previous_response[user] = 'A: {} B: {}'.format(text3, chat_response2)
else:
new_string = '{} A: {} B: {}'.format(previous_response[user], text3, chat_response2)
previous_response[user] = new_string
# Check if the string is over 1000 characters
if len(previous_response[user]) > 1000:
# Truncate the string to the last Query characters
index = previous_response[user].rfind('A:')
previous_response[user] = previous_response[user][index:]
print(previous_response)
@rare crag
rare crag
Hmmm, I'd consider doing your character check while building the dict, that way you can decide to truncate an entire message, rather than just chopping off the dict at an arbitrary length and sending some data from a a partial response that doesn't make much sense
rare crag
ahh ok, I misread that last part
oak wedge
index = previous_response[user].rfind('A:')
rare crag
Just be sure to include somewhere in the final prompt that the messages prior to the new one are part of the original conversation. Otherwise it may respond with something that considers the whole thing as one new prompt instead of a history + new prompt.
oak wedge
im using this:
query2 = "{}\n The text above is from the previous conversation, please refer to it when responding to the "
"next sentence below. \n {} do not include \n in your response".format(previous_response[user], query)
oak wedge
its already in prod. twitch.tv/all_in_89 Thanks for the suggestion!
oak wedge
thanks!
rare crag
Hmm, it thinks I'm it's new boss apparently xD
oak wedge
Updated with new prompts for better interaction as well as the ability to add to multiple channels.