Hey! I got interested in poking into this again with all the new OpenAI models coming out, so I made some code changes that you might be interested in looking at. Have a look at the commit pushed to the fork [https://github.com/markphaedrus/OpenCharacters]() . Summary of the changes: * Updated the list of OpenAI broadly-available models. * Model selector in the character editing page, and model selector in the thread window, now include pricing info. A little warning icon is displayed next to the model selector in the thread window if the pricing data is too stale. * Updated the default model choices to reflect what seemed like the most plausible choices in the current model set. * Added error handling for cases where a thread is using models that aren't supported anymore. * Tweaked the embedding code so that the subtraction from the calculated distances is model-specific, since the newer and cheaper text-embedding-3-small gives a very different range of values than what text-embedding-ada-002 did. * Added logic to avoid feeding temperature selections into models that don't support them (some of the new models don't, and throw errors if you try). * Likewise, added logic to change max_tokens to max_completion_tokens on a model-by-model basis, because some models don't support max_tokens and cry if you try to use it. * Updated the default characters so they use the current "good" model instead of a hardcoded choice. * Removed the gpt-4-specific "can we run this model?" check, since it didn't seem to be working when I tried to apply it to the current still-in-testing model. * Token-counting just uses the GPT-3 tokenizer for all OpenAPI models, because honestly I wouldn't even know where to begin updating the tokenizer. * Added code to let models specify an instruction-to-AI role other than "system", because at first it seemed like some of the new models didn't behave correctly unless I changed the role from "system" to "developer". That turned out to apparently not be true, so everything still uses "system", but I left the code in place just in case. * Added a "/flushembeds" command, to clear out the text embedding cache. Updated the tooltip in the thread input box.
This issue appears to be discussing a feature request or bug report related to the repository. Based on the content, it seems to be still under discussion. The issue was opened by markphaedrus and has received 1 comments.