Hexa PA has been updated.

• v0.3.0 is a major update, that comes with fixes after an API update broke it, available models updated to the latest gpt-3.5-turbo-0125 and gpt-4-turbo-preview, and the context management has been bufeed.

• When set to GPT 3.5 Turbo, Max Input Tokens = 0 (unlimited), Max Context Messages > 0, Max Output Tokens > 0 it switches seemlessly between 4k(latest) and 16k(legacy) models depending on how long the messages are.

• It's no longer an even split between input an output tokens, it can now take advantage of GPT-4 Turbo's 128K context even though it only does 4k output. (Successfully tested with more then 4k context.)

Since my last update, I implemented: - OpenAI came out with new models on 13th of June and HexaPA now uses the latest GPT-3.5 model, and it's new 16K variant. You can switch to longer context by defining higher then 2048 token limit, up to 8192. (HexaPA only does even split between input and output tokens for now...) - Reference injection - injects a reference to the rules into the prompt (only the prompt but not into the context) that makes the AI a bit more likely to follow the rules. It can be defined on the rules sceen. - The rules can now be exported to JSON, but not used as preset yet. (This feature is not finished, I just wanted to release an update with the new models, and longer context lenght ASAP.) - I also set up a github project page for HexaPA development plans, you can now see real time what features I'm working on, what is planned, and what is done arlready: https://github.com/users/RPBCACUEAIIBH/projects/7/views/1

As always you can download and test HexaPA from here: https://github.com/RPBCACUEAIIBH/HexaPA

Update

• Since my last update, I have renamed the project to HexaPA, due to pending trademark of the abbreviation "GPT" by OpenAI. PA suffix is a better name for it anyway, whether it stands for "Productivity Accelerator" (which is the whole purpose of this software) or "Personal Assistant".
• I added tooltips that show up at the bottom of the screen when the mouse hovers over a button, entry or textbox.
• I also added export feature, which exports it to JSON for now. I plan to add CSV, and a simple plain text support, as well. (JSON is vital for those who may want to preserve conversations when updating...) The last 2 images shows what rules and context ware sent to the AI in the debug output, and also that the export contains the block IDs for the same rules and context for the prompt. (The highlighted stuff on both images.) Thus the exported JSON can be used to fully reconstruct and/or analyze the conversation. This may be essential if it is ever used for training data, or to investigate why the AI responded a certain way since the context can be cherry picked by the user, so it's no longer just a linear conversation...
• I updated the token counter to work with generated context, as it was still not counting the context tokens.
• I also restructured some code, which is not a visible change, but vital for maintenance.
• I fixed a token limit bug. (It was limiting individual messages to the set value not the entire rules + context + prompt.)
• And finally published the first pre-release, since it is now a fairly usable AI chat application like ChatGPT.

Available on github: https://github.com/RPBCACUEAIIBH/HexaPA

Edit: The project name was changed to HexaPA to comply with OpenAI's branding recommendations.

• HexaPA is now on github, you can try it at: https://github.com/RPBCACUEAIIBH/HexaPA

Update:

• You can now enter a user name at first start. (It's still a single user application for now.)
• I've implemented automatic and selective context generation.
  • In Auto Context mode it includes the last n number of messages provided that it doesn't exceed the allowed token count.
  • If Auto Context is disabled it only includes the last n number of explicitly included messages unless it exceeds the allowed tokens.
• You can set the max number of messages you want to include as well as the allowed tokens in the Main screen.
• I have also added an Uniclude All button so you don't have to scroll back and individually uninclude messages if you want to change the list of included messages.

Some screenshots of my current AI project I've been doing in the last 3 weeks.

What is it? A productivity oriented, cross platform, desktop GUI application, much like ChatGPT, (using the same AI) but with privacy and full control over the language model.

The plan for HexaGPT v1.0: - Make it secure. (Done, it can handle files locally with reasonable encryption and compression.) - Functional as a chat application. (Almost there, it really needs to send some context to the AI, and a button to expand/contract prompts and responses...) - More control over the AI then ChatGPT allows. (To be implemented...) - Rules (Which define the AI's personality. Implemented, but I'd also like to add some presets for general use.) - Automatic context limited by token allowance set by the user (not yet implemented) - Selectable context (not yet implemented) - Automatic recursive context (Maybe! I had that idea today while playing with rules... Not sure how viable it is, I'll have to experiment, but it could allow HexaGPT to "remember" something from the very beginning of the conversation. If enough tokens given for several successive API calls with potentially large amount of context...) - Adjustable AI parameters (Not yet implemented, but it isn't hard, it's a matter of time.) - Since I'm a Linux only user since 2011, and not afraid of the teminal, it must have CLI interface as well. :) (Partially implemented...) - I also plan support for Open Assistant, and Bard, possibly others in the future, but currently none of those have publicly available APIs, so maybe not for v1.0

The software will be free and open source of course(as the BSD 3 clause license indicates), but it is using GPT 3.5 Turbo model via OpenAI's payed API so you must pay OpenAI to use it. (Although it's cheap for text based applications at the moment.) I've already requested access to GPT4, still waiting on their response.

Coming soon: It isn't yet a usable chat application like ChatGPT(but it's slowly getting there), it only sends the prompt and rules to the AI so far, I've just added the entire "rules" part in the last 2 days. It doesn't send any context to the AI of the previous prompts and responses so the AI seems to have "forgot" what you've told it seconds ago and what it responded. To be honest I do have a bad habit of abandoning half complete projects when I can't figure out something, but with the help of ChatGPT I have high hopes for this, especially cause I do it for increasing my own productivity.

P.S. Since I posted this over on Patreon(a few days ago), I also made all the textboxes in the chat screen flexible, so now if you click on a textbox it expands to a larger size while all others shrink. Added a few more hotkeys, and did some essential but not so visible code changes in order to implement individually selectable context, although I'm still working on that feature.