Hey @jackschedel !

Thanks for the awesome repo!

Azure Openai models are named as 'gpt-35-turbo-1106' instead of 'gpt-3.5-turbo-1106'. Could you add a mapping for azure endpoints by checking the endpoint URL? Azure endpoints are in this format: 'https://DEPLOYMENT-NAME.openai.azure.com/'.

Something like this would work: if 'openai.azure.com' in endpoint_url, change 'gpt-3.5-turbo-1106' to 'gpt-35-turbo-1106'.

Bests

the other software named "better chatgpt" is ok and will not pop error

I build the latest version of KoalaClient , which pop this alert.

after upgrade to latest, have to delete all config and restart the software. The old config files have to be exported and imported
imported error

• spinwait icon change on processing
• auto-stop and wait for process end on "save and submit"
• separate record+submit button
• option to auto-stop after 5s of silence (slider?)
• add usage to API cost

Is anybody having bugs like:

• 
  With settings such as:
• 

This is annoying and I will take a look at it soon 😋

may need to come with full rework of chat dragging behavior (preventDefault and have custom animation and behavior)

i.e. I could specify that T-binding is always changed to keybinding, toad to code etc

would start as electron only, implementing local temp files on website sounds like a pain

If I try to move my cursor to a previous word by using Ctrl+Arrow key, it changes documents/chats?? Please don't hijack basic keyboard shortcuts.

If the timing is just right, the "Stop Generating" button can cause some wacky behavior - such as two responses being written at once to the same message, or the response stream getting appended to the last user message (rather than a new assistant message).

Usually this happens when I quickly edit and re-send a message (in under a second).

I think the reproduction trigger is pressing the "Stop Generating" button after the request is sent, but before a response stream has been received.

A simple fix might be to not show the button until the response has been received, although it might be better to also abort the request with the button.

Has anyone experienced this, and am I understanding correctly?

There are a bunch of models on https://openrouter.ai/docs#models, some of which are free. Is it possible to support them?

Do you have any plans on integrating Claude API not just OpenAI in your fork?

Thank you.

The error message in the API popup is not visible due to a color contrast problem.

It can be fixed by making the text color to white.

A couple small things to be ironed out with the max content setting calculations.

max_token calculation

If I'm understanding this correctly, the API considers max_tokens to be the context and the response combined. So there shouldn't be a need to subtract the max context setting from the max token setting.

Currently we are calculating the parameter like this:
max_tokens = (maxTokenSetting - maxContextSetting)

We should simply do this before each request:

• set the maxTokenSetting directly as the parameter in the request
• trim the messages by the maxContextSetting

^{Or, to calculate on the fly similar to this, we'd have to introduce a third setting, like I described in another issue:
max_tokens = (maxContextSettings - desiredMaxResponseSetting) (with the maxTokenSetting still serving to cap this at a maximum)}

Other likely manifestations of this bug:

• Sometimes requests are sent with a max_token value that is too large to be allowed:

(response is an error because the maximum value for max_token on gpt4 is ~8000).

Feature Request

Summary:
Allow users to input custom model names in the model selection dropdown or within the settings when choosing which AI model to interact with.

Problem:
Certain proxies for OpenAI services have been experiencing issues where the /models endpoint might be malfunctioning or is not populating the model selection list as expected. This results in users being unable to select from the full range of available models, potentially limiting their access to specific or preferred AI models that are otherwise operational.

Proposed Solution:
Implement a feature within the chat interface that enables a user to add custom model names directly to the model selection dropdown. This could take the form of an input box within the dropdown itself, allowing for direct typing of the model name. Alternatively, a small section in the settings could be dedicated to managing custom model names where users could add or remove model names that they frequently use or wish to test.

Benefits:

• Provides a workaround for when the model list is not automatically populated.
• Offers flexibility to advanced users who are aware of model names that are not listed but are available for use.
• Ensures continuity of user experience by enabling access to all models, regardless of potential temporary proxy endpoint issues.

i.e. neovim extension, chrome extension

I've had an idea for a while now:

Similar to the 'Model Swap' button you've added, I also frequently adjust temperature, top_p, and _ presence penalty_ from my defaults l, according to my needs at the time.

For example:

• when referencing an obscure programming language, we want accurate and deterministic chat settings.
• when writing a short story, accuracy becomes less important, but we want a creative chat.

Perhaps:

• the user could set up a few configurations of settings
• to quickly switch between them (like the Model Swap button in this fork)
• eventually each setting option could even have checkbox to omit changing from the default?

happens when selecting and copying text using ctrl-a or dragging below the final line

Sometimes the max_token parameter for the title generation is negative:

The response is something like:

{
  "error": {
    "message": "-228 is less than the minimum of 1 - 'max_tokens'",
    "type": "invalid_request_error",
    "param": null,
    "code": null
  }
}

And other time it is way too long, and I get titles that are an entire paragraph.

Just notice that the "Max-Tokens" API setting isn't getting included in the chat request body, even though it appears to be a user configuration setting, and is likely being stored.

• Max Tokens parameter is a control for the maximum number of tokens that can be generated in a single call to the GPT model

Automatically set model according to settings & chat size

Proposed feature:

This would be best implemented as an optional setting, that a user could opt into. If enabled, the only models to choose from in the chat config would be:

• gpt-3.5-turbo
• gpt-4

The token length models could then be automatically selected only when required, determined by these factors:

• Chat config settings:
  • Max tokens
  • Max context
• The size of the message ^{(for example, it is not necessary to use the gpt-3.5-turbo-16k model if you are well under the 4000 token limit for gpt-3.5-turbo-16k)}

likely electron only because system clipboard access on web is weird.

SHIFT + ENTER behavior when editing existing chat messages.

I am frequently editing existing System/User/Assistant chat messages. It is a useful technique to:

• establish context
• reduce API token usage (cost)
• "rewind" the conversation to a previous state, perhaps after a tangent

Notice that:

• When composing a NEW message:

  • Pressing just the ENTER key sends the message
  • SHIFT + ENTER creates a new line without sending the message.
    • this is common already (e.g. Facebook Messenger, etc.), and many users will already intuitively understand it.

• However, when editing an EXISTING message:

  • Pressing just the ENTER key creates a new line
  • SHIFT + ENTER will send the message

Because SHIFT + ENTER explicitly sends the HTTP request, it seems like this might've been deliberate by the original author? But I can't see any reason as to why!

I've even been using a UserScript to accomplish this for a few weeks now, so I feel familiar with each UX.

But if there is a good reason for the inconsistency, it could be implemented as something that the user could toggle on or off, in the configuration settings.

https://stackoverflow.com/questions/31749625/make-a-link-from-electron-open-in-browser

(discord has this keybinding)

maybe down arrow to edit most recent Assistant Chat?

maybe want to add a setting toggle to this, and/or have different behavior for the left menu button and the mobile bar button

(Whisper is currently disabled on the website because it doesn't work properly. Remove the isElectron() && on line 266 of EditView.tsx to enable it on the website.)

Electron:

Website:

Steps to Reproduce:

• Click on the "New Chat" button without sending any messages in the existing chat.
• Observe that the existing chat is renamed to "New Chat 1."
• Repeat steps 2 and 3, and note that subsequent existing chats are renamed incrementally (e.g., "New Chat 2," "New Chat 3," etc.).
• Rename the chat to anything and send a message
• Click on the "New Chat"

is it ok to support this ?

gpt-4-0125-preview is a new api added in Openai.
we can add it.
Thanks!!

will likely want line numbers in conjunction with this