shipbit / slickgpt Goto Github PK

When i try to use gpt-4 or gpt-4-32k it says model does not exist when sending the query

Title is self-explanatory. I wish ChatGPT had this, so I'll add it to this project. But there should be a command palette or a search bar that lets you search through all your chats, this would be very useful. I'll work on an initial PR.

SlickGPT users should be able to cancel the generation of completions like on chat.openai.com:

Hello! This project is very 👍
I want to know how to deploy it with vercel.
I fork your project, and create a project with vercel. but when I click deploy, there was many errors.

This is my deploy setting:

This is the error informations:

[22:57:21.186] Running build in San Francisco, USA (West) – sfo1
[22:57:21.257] Cloning github.com/ddy-ddy/my-personal-chatgpt (Branch: master, Commit: 1b19c0f)
[22:57:21.389] Previous build cache not available
[22:57:22.355] Cloning completed: 1.098s
[22:57:22.480] Running "vercel build"
[22:57:22.913] Vercel CLI 28.18.5
[22:57:23.468] Installing dependencies...
[22:57:36.361] 
[22:57:36.361] added 411 packages in 13s
[22:57:36.361] 
[22:57:36.361] 34 packages are looking for funding
[22:57:36.361]   run `npm fund` for details
[22:57:36.387] Detected `package-lock.json` generated by npm 7+...
[22:57:36.388] Running "npm run build"
[22:57:36.678] 
[22:57:36.679] > @shipbit/[email protected] build
[22:57:36.679] > vite build
[22:57:36.679] 
[22:57:37.536] 
[22:57:37.541] �[36mvite v4.2.1 �[32mbuilding SSR bundle for production...�[36m�[39m
[22:57:37.576] transforming...
[22:57:39.127] 
[22:57:39.127] �[1m�[33mwarn�[39m�[22m - As of Tailwind CSS v3.3, the `@tailwindcss/line-clamp` plugin is now included by default.
[22:57:39.127] �[1m�[33mwarn�[39m�[22m - Remove it from the `plugins` array in your configuration to eliminate this warning.
[22:57:43.242] �[32m✓�[39m 831 modules transformed.
[22:57:43.259] 2:57:43 PM [vite-plugin-svelte] ssr compile done.
[22:57:43.259] package          	files	  time	   avg
[22:57:43.259] @shipbit/slickgpt	  645	 1.62s	 2.5ms
[22:57:43.259] @sveltejs/kit    	    1	24.4ms	24.4ms
[22:57:43.260] �[32m✓�[39m built in 5.72s
[22:57:43.260] �[31m"FIREBASE_APIKEY" is not exported by "�$env/static/private", imported by "src/misc/firebase.ts".�[39m
[22:57:43.260] file: �[36m/vercel/path0/src/misc/firebase.ts:4:1�[39m
[22:57:43.260] �[33m2: import { getDatabase, get, ref } from 'firebase/database';
[22:57:43.260] 3: import {
[22:57:43.260] 4:   FIREBASE_APIKEY,
[22:57:43.260]      ^
[22:57:43.260] 5:   FIREBASE_AUTHDOMAIN,
[22:57:43.260] 6:   FIREBASE_PROJECTID,�[39m
[22:57:43.262] �[31merror during build:
[22:57:43.263] RollupError: "FIREBASE_APIKEY" is not exported by "�$env/static/private", imported by "src/misc/firebase.ts".
[22:57:43.263]     at error (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:2105:30)
[22:57:43.263]     at Module.error (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:13174:16)
[22:57:43.263]     at Module.traceVariable (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:13559:29)
[22:57:43.263]     at ModuleScope.findVariable (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:12061:39)
[22:57:43.263]     at Identifier.bind (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:7933:40)
[22:57:43.263]     at CallExpression.bind (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:5722:28)
[22:57:43.263]     at CallExpression.bind (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:9470:15)
[22:57:43.263]     at ExpressionStatement.bind (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:5726:23)
[22:57:43.264]     at Module.bindReferences (file:///vercel/path0/node_modules/rollup/dist/es/shared/node-entry.js:13170:18)�[39m
[22:57:43.294] Error: Command "npm run build" exited with 1
[22:57:43.657] Deployment completed
[22:57:43.613] BUILD_UTILS_SPAWN_1: Command "npm run build" exited with 1

Apparently everybody uses dark mode ;)

Hi,
I have made some attempts on https://slickgpt.vercel.app but without a correct Api key

but it looks like my chats with my APIkey are being redirected to OpenAI through the website.
Wouldn't it be safer if my browser sent the requests directly to OpenAI and not through your server? At chatworm.com it works exactly like this.

Regards,
Theo

When GPT is done typing the text box is disappearing completely and then reappearing.

Slow-mo capture:

Integrate the WhisperAI API to let SlickGPT users input prompts with their voice. This would be especially nice on mobile.

The app works fine locally, but when I run it on my server, and then I access vps-ip-address:5173/ the app doesn't load at all.

I'm no expert in node or npm, what I am missing here ?

Numbered lists don't go higher than 1. Also spacing between is off.

Example:

When users have multiple devices or are about to retire old devices, they prefer to transfer data from one to another before deprecating old ones. It would be helpful if a dialogue import/export mechanism or tutorial could be provided, or one step further, a syncing mechanism between various devices of a single user.

On chat.openai.com you can edit user prompts which "branches" the chat:

• All prompts and completions after the edited one are still accessible and can be switched in the UI
• The edited prompt creates a branch in the chat
• Whenever you continue prompting in a branch, only the prompts and completions in your current branch are sent to the ChatGPT API (to save tokens and keep a clean context). Our token cost estimation has to reflect this correctly.

SlickGPT is setup to use Vercel as hosting provider. It uses @sveltejs/adapter-vercel and configures the endpoints in the /api dir to run as Edge functions. Create some documentation and/or branches to show how to run it on:

• Netlify
• Azure
• ?

I already tried with Netlify and encountered an unexpected issue where Netlify was not able to access env vars from $env/static/private. See this issue for more infos and a workaround. Another problem was that longer prompts timed out because the edge/serverless functions ran too long on Netlify.

When setting the dialogue title manually, pressing enter key should trigger "Save" (finish setting the title), rather than Automatic suggestion via GPT3.5 .

The current version is shown below:

I think the cryptr module on NPM would be an excellent option for this. There'd need to be an extra env variable for storing the serverside key, the FIREBASE_APIKEY variable could be a fallback since all that is needed is a random string, but I think it would make the software seem more secure. I can work on a PR to add this

I have not been able to find what's applying a height: 20px inline on the textarea, but it's breaking things.
Firefox, latest, mobile and desktop.

Expectation

Reality

In Chrome it seems to work fine, and a height of 42px is assigned (desktop), so I'm guessing there is some js or post processor that I don't know about doing it, so I'm posting here.

It would be nice of custom assistants were supported https://platform.openai.com/docs/assistants/overview

Personally I am a digital pack-rat. I like to keep my history of chatgpt chats from openai.com. It would be nice to import my previous chat history into this interface.

I also use the cli application aichat https://github.com/sigoden/aichat that I would like to import if possible. They store everything in a single .md file though so I think that would be a lot more difficult.

See screenshot. I wonder if there's a way to close any open backticks temporarily while streaming in the response.

Is there any plans currently to support the new GPT-4-vision model in the future for uploading images in chats?

Hi,
I tried sending a SVG element in a message, but it was actually rendered, making the page +/- unreadable
Also some very long lines were ellisped in the message I got back
Maybe we need horizontal scroll + escaping input (and output)
Thanks

Would be nice to not need to copy and paste the context from a previous thread (if you want to use the same context, but in a different thread).

Maybe SlickGPT could provide a library of context "templates".

I'm trying to select text, copy it and pas to my editor....
It is impossible for some reason... The selection seems not to work.
It is the same for the frame where i enter my text and for frames with my previous input, GPT output too, and context editing frame too,

I'm running Chromium Version 111.0.5563.146 (Official Build) snap (64-bit)
Under Ubuntu 18.04

Editing an existing message (①) should create a new branch first, and then the reply can be generated under the new branch like (②).

In the current version, editing a message and request completion will result in the reply being generated under the current branch rather than the new one, perceived as the chat not being branched properly (③), then the reply is moved to the new branch after generation.

Hide or make it unreadable, once the API key is entered. Makes it easier to record videos or show this to someone else without spoiling the key ;-)

streaming response not working in Iphone 7 IOS 15.7.3 - safari browser.

it may work with json reponse.

i have tried to test it using my own python code. it ndont work either.
@app.post('/chatgpt-proxy')
async def chatgpt_proxy(request: Request):
try:
data = await request.json()
except json.JSONDecodeError:
raise HTTPException(status_code=400, detail="Invalid or empty JSON request body")

headers = {
'Authorization': f'Bearer {OPENAI_API_KEY}',
'Content-Type': 'application/json',
}
print(data)
print(headers)
r = requests.post(OPENAI_API_URL, headers=headers, json=data, stream=True)
return StreamingResponse(r.iter_content(chunk_size=8192), status_code=r.status_code)

The share feature in SlickGPT requires a database because we somehow have to pass the Chat object from client A to client B.

The project is currently set up to use a Firebase Realtime database as easy "JSON dump storage".

Create a branch and some documentation where the Share feature can easily be disabled and all Firebase dependencies (mostly the env vars) are removed from SlickGPT so that devs can run their own instance quicker and with less hassle in case they don't need to share chats.

Also document more clearly which code devs would have to edit to use another database provider than Firebase (which is basically just the share endpoint)

When the output is recognized as JSON it is formatted as code using <pre> and <code>.

But when the JSON contains a nested object, the top-level isn't formatted as code anymore.

examples:
https://slickgpt.vercel.app/shared/acidic-shallow-actor

Currently the Max Tokens value has a maximum limit to 128000 for model gpt-4-1106-preview, however, from OpenAI documentation , the model returns a maximum of 4096 tokens, making the most of the dragging bar useless. Setting the values to exceed 4096 will result in an error in requesting a reply.

I think it would be great to have docker images ready. It might make deployment easy and quick.

Whenever I send some input and put my device in standby or the app in background while the answer is being written, both my input as well as the GPT answer get erased from chat history.

Likely a bug?

I'd prefer my input wouldn't be lost and even if the response is not complete, that the received data remains as well.

iOS 16.4.1 (a) - iPhone 13 Pro Max

Hi, I managed to run slickgpt on my own server, using @sveltejs/adapter-node, and behind nginx proxy with Let's Encrypt SSL.

When I try to use SuggestTitleModal.svelte I get some errors. They only show when I connect via my domain. If going through IP:port directly, all works fine.

JSON:

Uncaught (in promise) SyntaxError: JSON.parse: unexpected character at line 1 column 1 of the JSON data SuggestTitleModal.svelte:573:2
    handleSuggestTitle SuggestTitleModal.svelte:573
    AsyncFunctionThrow self-hosted:856
    (Async: async)
    listen index.mjs:463
    listen_dev index.mjs:2325
    mount SuggestTitleModal.svelte:122
    mount SuggestTitleModal.svelte:459
    m svelte-hooks.js:291
    mount_component index.mjs:2097
    mount @skeletonlabs_skeleton.js:25390
    mount @skeletonlabs_skeleton.js:26472
    mount @skeletonlabs_skeleton.js:25276
    update @skeletonlabs_skeleton.js:26709
    update index.mjs:1347
    flush index.mjs:1307
    (Async: promise callback)
    schedule_update index.mjs:1258
    make_dirty index.mjs:2133
    ctx index.mjs:2171
    instance32 @skeletonlabs_skeleton.js:26759
    set index.mjs:34
    update index.mjs:42
    trigger stores.js:10
    showModalComponent shared.ts:129
    handleEditTitle Toolbar.svelte:13
    (Async: EventListener.handleEvent)
    listen index.mjs:463
    listen_dev index.mjs:2325
    mount Toolbar.svelte:80
    mount Toolbar.svelte:203
    m svelte-hooks.js:291
    mount_component index.mjs:2097
    mount +page.svelte:132
    mount +page.svelte:1315
    m svelte-hooks.js:291
    mount_component index.mjs:2097
    update root.svelte:292
    update_slot_base index.mjs:100
    update +layout.svelte:89
    update_slot_base index.mjs:100
    update @skeletonlabs_skeleton.js:5920
    update index.mjs:1347
    flush index.mjs:1307
    (Async: promise callback)
    schedule_update index.mjs:1258
    make_dirty index.mjs:2133
    ctx index.mjs:2171
    $$set root.svelte:619
    get proxy.js:83
    $set index.mjs:2272
    key proxy.js:46
    navigate client.js:1096
    InterpretGeneratorResume self-hosted:1455
    AsyncFunctionNext self-hosted:852
    (Async: async)
    _start_router client.js:1569
    (Async: EventListener.handleEvent)
    _start_router client.js:1485
    start start.js:27
    <anonymous> (index):30
    (Async: promise callback)
    <anonymous> (index):29

also in GET:

NS_ERROR_WEBSOCKET_CONNECTION_REFUSED

and in POST:

Status
403
Forbidden
VersionHTTP/2
Transferred405 B (46 B size)
Referrer Policystrict-origin-when-cross-origin
Request PriorityHighest
DNS ResolutionSystem

It would be cool if SlickGPT users could see the latest changes within the app. Maybe on the Dashboard?
UI and technical concept needed - where would these infos come from and where would we display them?

Reported by several users:
Typing on mobile or slow clients leads to a noticeable delay. This is probably happening as a result of the live token cost calculation.

Dark Mode

Light mode

Browser: Microsoft Edge Latest

I am currently using GPT-4 to generate responses to my requests. However, as the generation process is slow, I find myself staring at a blinking textbox for minutes at a time. As a result, I am unable to read the output that has already been generated.

Implement the Skeleton UI theme switcher seen on their page: https://www.skeleton.dev/.
If this is possible, please save this setting in local storage as well, so I don't have to change it everytime.

slick indeed!
Would be cool if you could add a copy code button when answer contains created code.

If you scroll chat boxes out of about two screens long, and bring them back in, the tooltip text Approximate Token Cost will appear very briefly, flickering when you hover over the tokens.

This will in many cases break the formatting of the partial message but is desired nevertheless. Users can always delete cancelled messages if they're bothering them.

Quote OpenAI:

The moderation endpoint is a tool you can use to check whether content complies with OpenAI's usage policies. Developers can thus identify content that our usage policies prohibits and take action, for instance by filtering it.

See OpenAI Docs

This endpoint is free and it would be nice if SlickGPT would integrate this API.

Would you please add other models like davinci and curie?

It would be nice to not have to go in and change from 3.5-turbo to 4 every time I make a new thread.

(Great app btw!)

chat.openai.org has this neat feature where they let the AI set a title for your chat based on the first prompt und completion in the chat. People are bad at naming things ;)

Condition: if (chat.slug === chat.title) and there is at least one question and answer in the chat

Ideas:

• either do this automatically as soon as the first answer arrived OR
• Ask the user to name the chat or offer a button to let ChatGPT pick one. Do this when the user closes a chat and above conditions are met. This could save some tokens. Bonus: Estimate how many tokens this will roughly cost.

Technical:
it probably makes sense to set stream: false for this (and other) calls to the OpenAI API. This would require passing stream as param to the the endpoint in the Ask API and adding a new handler in the client that receives the answer. Currently stream is always hardcoded to true and the handler in ChatInput.svelte always treats the answer as EventStream.

It'd be nice to scroll up and view convo history while gpt is "typing". This is currently not possible because of auto scroll. It's more noticeable when using gpt4 because it's noticeably slower than older models (at the time of writing at least).

Possible solutions:

• (best) only auto scroll if the user is at the bottom of the convo
• manually toggle auto scroll on/off
• nothing, at the rate we're moving gpt4 speed will soon be a non-issue

Currently I cannot select/copy text during an ongoing response.
Issue? My selected text is immediately unselected.

Why would I want an incomplete selection?
Example: Big code request where the ai gives multiple code blocks and may have already finished one.

Why fix? This saves time as I would not need to wait for the response to finish before I can select and copy the text.

Would be very appreciated.

Browsers: msedge/117, chrome/114

In generating code-formatted long texts that do contain new lines, the response bubble won't wrap and exceeds the screen. Image below:

The properly wrapped format is demonstrated after generation is complete. Example:

SlickGPT uses gpt3-tokenizer which is one of the few libs I found that "just runs" in the browser. It's close enough but not perfect as the calculations in

Another and bigger problem is performance. The gpt3-tokenizer has a huge payload and it's pretty slow. Other solutions use advanced stuff like Node Buffer structures and are much faster. The problem is that they don't run easily in the browser without Node.

Any ideas how to calculate the tokens per Chat (context, prompts, completions) more accurately and faster?