ChatGPT ConvoBot - ready-to-go bot interface for superb AI conversation! Now with Nanite Vox and unicode support! Version 1.7 December 3 2023

Bot/AI/GPT Blog: https://johnsonbots.blogspot.com/

This new system allows you to have conversations with the amazing new ChatGPT conversation bots from OpenAI. This system handles rather deep-ply conversations with memory, and is not simply a question-answer system. The result is not just a chatbot, but is an amazingly human-like simulation of a real person.

The system requires you to have an OpenAI account (visit https://platform.openai.com/ ). This system currently uses gpt-3.5-turbo which is very cheap, efficient, and effective. It is possible that you can get free OpenAI access (there was a trial 3 month $5-value free access subscription in early 2023) but you will likely need to have a paid subscription. I find it to be very cheap for the fun I've had with it!

You can also change the model to gpt-4 if you have access to it but beware - 4 appears to be about 10x the price of 3.5!

If you want the thing to actually *talk* in audio, check out my blog article
https://johnsonbots.blogspot.com/2023/07/have-yourself-automatically-speak-in.html
The free code in there can easily be adapted to put the ConvoBot output into audio!

The system now also works with OpenAI-compatible APIs - make sure the endpoints end with /completions and are compatible with the gpt-3.5 and gpt-4 /completions endpoints!

The thing can be worn as an attachment, or rezzed as an in-world object. There is an API-like pair of scripts that will let you modify it to do what you want - e.g., accept input from some device or other, send output to some other device, via channel messaging, link to other object, etc. But as delivered it simply listens to nearby chat and then replies.

The system as delivered does not respond to "attachment-talk" or "green talk", but only to avatar "white chat". This can be altered easily by editing the i/o script.

The personality of the bot is defined by a "biography" that is written in plain English. There are a number of example bio's included, to illustrate using the bot as

1. a good friend named Sally (built in on delivery)
2. a store worker
3. a nanny
4. a wild and crazy party girlbot
5. a famous and annoying scientist

These can be altered and tuned to get the bot to behave how you like it. The results are really amazing! Enjoy the system and let me know if you set up a bio you really like - I'd love to add it (suitably edited for privacy) to the library of possibilities.

Finally, a version is included which works with the Nanite Systems vox system - output goes to vox and gets processed by its filters! The device is attached under the name GPTAI.

Version 1.7 adds a command to change the API web address (URL) to something other than OpenAI's. In mid-2023 a number of competing services set up compatible APIs and this allows you to access them by changing the URL used. A couple of customers have used the system with alternate API URLs.

GENERATIVE AI NEWS - Thanks to BonnieBelle86 for pointing out the openrouter system - another OpenAI-api-workalike! It works fine with my systems, just substitute the OpenAI URL with https://openrouter.ai/api/v1/chat/completions and get an account and API key from https://openrouter.ai/ , stick that into my code too. Lots of models available and at very competitive prices - see https://openrouter.ai/docs#models

Version 1.6 repairs a problem introduced by a format change in http return by the OpenAI API on 6/8/23 - please contact me if you have any garbled output issues. The fix is to replace the gptconnect-bot... script which requires rez at a sandbox or a place you have rez rights.

Version 1.5 adds "speaker id" so that GPT will know who said what. There are also filtering lines put into the gptconnect-io script to respond only when a "trigger" word is heard in input (e.g., name) and/or when the speaker is less than a certain distance away. Let me know if there are any problems!

Version 1.4 adds long-term memory of conversations that will survive reboots, and also provides user-defined filtering of out-of-character responses to make conversations even more realistic.

Version 1.3 improves memory usage so that the thing should run out of memory much less often.

Version 1.2 supports full unicode/utf-8 in and out despite the mishandling of it by the OpenAI API. Now you can ask it to speak any foreign language and even translate from one language to another...!

The slightly revised version 1.1 fixes a few small errors with, in fact, error handling.

You should be able to request redelivery to update your systems.

I also did some billing tests. The price of using this thing appears to be about 1000 tokens = 1k token per prompt-response (depends on the bio you use and what you ask!). At present this means billing from OpenAI at the level of $0.002 USD per prompt-response, or a penny for 5 prompt-responses. What else can you get for a penny these days?

Have fun!
Stone Johnson (musichero)
stone.johnson.78@gmail.com