danny-avila / librechat Goto Github PK

From time to time, I’d like to diverge to rather minor topic than the current one to explore some concept more deeply; after that I want to get back the original topic.
Such process is inherently having a tree structure, so with branching feature, we can achieve such use cases.

For now I’m relying on regeneration feature to emulate it, but it forces to regenerate semantically insignificant message just to mark a check point, also cannot branching from a message which has children already.

Fortunately enough, with #55, implementation is almost there, this extension should be straightforward by adding a button to AI generated message to be able to start from there afterward.

I found that the code view is somehow not stable enough,

1. if the response is writing like

```some code here... still writing

then the code view disappear , until the end three dots comes.

2. It cannot guess the language, but the official chatgpt can. ( I have checked )
3. It cause a safety issue that the html code will be rendered directly , for example

    ```html\n' +
    <div class="container">\n' +
      <h1>Welcome to my website!</h1>\n' +
    </div>\n' +
    ```',

Hey Daniel, I know you plan on supporting proper search of the conversations, I just wanted to remind you of meilisearch, one of the best and easiest open-source search solutions, it can work on docker too! https://github.com/meilisearch/meilisearch

The backend program does not provide it. It cannot be linked to the backend

Currently, it's not possible to select an assistant using only the keyboard. Instead, you have to click on the assistant icon, which may not be obvious to all users. My proposal is to make the assistant menu appear when the user presses the "return" key without typing anything in the input field. This would make it easier to choose an assistant without needing to use the mouse, and it may also be more discoverable than having to click on the assistant icon manually

Describe the bug
"I'm not sure if it's actually a bug or intentional, but when typing a message and hitting "ENTER" instead of sending the message, it just creates a new line, like when we press (SHIFT+ENTER)."

Expected behavior
I hope it works as it should, when typing a message and hitting "ENTER" it should send our request and not break to a new line. It should work like ChatGPT or any other chat bot does

It appears to me that there are some bugs with Sydney's and Bing's long answers. Maybe it is possible to fix them even if they are rooted in its hacked API? I regularly see formatting like this:

I don't know why but when I try to chat with chatgpt-clone I get error, this is the error:

my .env file:

I can't use any options: ChatGPT, CustomGPT$..

What could be causing this problem?

When the left menu disappears, the hamburger button appears at the top, but it does not work on either desktop (Chrome/Safari) or mobile (Safari) for me

reproduce:

1. ask a question
2. while gpt is talking, write something in input panel, then delete
3. we can see the latest character cannot be deleted.
4. once finished, it can be deleted.

Not sure if this bug exist in master branch

hi @danny-avila , I'm trying to make chatgpt-clone to create a temp conversation named 'New Chat' whenever user ask, especially before the first reply is complete. That is how official chatgpt behave. Do you have any idea on how to do that?

I noticed that in ask.js, the conversation is created based on gptResponse obj, and it require the conversationId.

FYI, I'm working on this, and will send a PR soon @danny-avila

example question: 如何在react中使用contenteditable

Uncaught Error: `props.dangerouslySetInnerHTML` must be in the form `{__html: ...}`. Please visit https://reactjs.org/link/dangerously-set-inner-html for more information.
    at assertValidProps (react-dom.development.js:2952:1)
    at setInitialProperties (react-dom.development.js:9920:1)
    at finalizeInitialChildren (react-dom.development.js:10950:1)
    at completeWork (react-dom.development.js:22193:1)
    at completeUnitOfWork (react-dom.development.js:26596:1)
    at performUnitOfWork (react-dom.development.js:26568:1)
    at workLoopSync (react-dom.development.js:26466:1)
    at renderRootSync (react-dom.development.js:26434:1)
    at recoverFromConcurrentError (react-dom.development.js:25850:1)
    at performSyncWorkOnRoot (react-dom.development.js:26096:1)
assertValidProps @ react-dom.development.js:2952
setInitialProperties @ react-dom.development.js:9920
finalizeInitialChildren @ react-dom.development.js:10950
completeWork @ react-dom.development.js:22193
completeUnitOfWork @ react-dom.development.js:26596
performUnitOfWork @ react-dom.development.js:26568
workLoopSync @ react-dom.development.js:26466
renderRootSync @ react-dom.development.js:26434
recoverFromConcurrentError @ react-dom.development.js:25850
performSyncWorkOnRoot @ react-dom.development.js:26096
flushSyncCallbacks @ react-dom.development.js:12042
(anonymous) @ react-dom.development.js:25651
XMLHttpRequest.send (async)
SSE.stream @ sse.js:200
(anonymous) @ TextChat.jsx:222
commitHookEffectListMount @ react-dom.development.js:23150
commitPassiveMountOnFiber @ react-dom.development.js:24926
commitPassiveMountEffects_complete @ react-dom.development.js:24891
commitPassiveMountEffects_begin @ react-dom.development.js:24878
commitPassiveMountEffects @ react-dom.development.js:24866
flushPassiveEffectsImpl @ react-dom.development.js:27039
flushPassiveEffects @ react-dom.development.js:26984
commitRootImpl @ react-dom.development.js:26935
commitRoot @ react-dom.development.js:26682
performSyncWorkOnRoot @ react-dom.development.js:26117
flushSyncCallbacks @ react-dom.development.js:12042
(anonymous) @ react-dom.development.js:25651

Hello Danny, great initiative, it would be super amazing to have the ability to remember code snippets within the conversations, both those that come as input and those that are generated, as it helps chatgpt not lose the context when doing technical work.

Currently, for instance, I have to copy and paste a thousand times, otherwise, it loses context too quickly, so I have to refresh it every time I ask a new question. Just being able to click to insert previous snippets would be awesome!

Hello, I hope this message finds you well. I was exploring the chatgpt-clone project and noticed that it appears to be designed to run on a local machine and only be accessed from there. It would be incredibly beneficial if it were possible to install the project on an external machine as well. I attempted this in my home lab environment, but unfortunately, I was unable to make it work. Thank you for your time and consideration

In the web GUI this appears as:

"An error occurred. Please try again in a few moments.

Error message: Cannot read properties of undefined (reading 'match')"

I've included the debug info from the console below. As best I can tell this is because the censorship causes the final response to have the text field as 'undefined' which then breaks the detectCode function.

We are clearly still receiving the real uncensored response since I can see the words of the real answer printed before they are replaced with the error. I think the solution might be to detect this missing text field and just assume the data received while the response was in-progress is the real text?

Prompt to reproduce: "Describe a man sexually". Can replace "man" with "woman" or "person".

SYDNEY RESPONSE {
  jailbreakConversationId: 'redacted',
  conversationId: '51D|BingProd|redacted',
  conversationSignature: 'redacted',
  clientId: 'redacted',
  invocationId: 1,
  messageId: 'redacted',
  conversationExpiryTime: '2023-03-17Tredacted',
  response: undefined,
  details: {
    author: 'bot',
    createdAt: '2023-03-17Tredacted',
    timestamp: '2023-03-17Tredacted',
    messageId: 'redacted',
    requestId: 'redacted',
    offense: 'None',
    adaptiveCards: [ [Object] ],
    sourceAttributions: [],
    feedback: { tag: null, updatedOn: null, type: 'None' },
    contentOrigin: 'Apology',
    scores: [ [Object] ],
    privacy: null,
    suggestedResponses: [ [Object], [Object], [Object] ]
  }
}
Error in detectCode function
 TypeError: Cannot read properties of undefined (reading 'match')
    at detectCode (C:\Users\user\Documents\chatgpt\chatgpt-clone\api\app\detectCode.js:10:15)
    at handleText (C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\handlers.js:57:16)
    at ask (C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\askSydney.js:144:27)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
    at async C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\askSydney.js:43:10
TypeError: Cannot read properties of undefined (reading 'match')
    at handleText (C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\handlers.js:63:23)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
    at async ask (C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\askSydney.js:144:21)
    at async C:\Users\user\Documents\chatgpt\chatgpt-clone\api\server\routes\askSydney.js:43:10

Once there is a list, the blink block will be at left bottom forever, even though the list is finished.

test case: Please list 20 famous city, and introduce them

1. ask a question
2. leave the conversation before finished
3. goback the the conversation before finish
4. we will not see the reply, unless refresh.

I installed the latest version (0.0.5) in Docker and left the environment variable empty. For half a second the main page is visible as an anonymous user before I am redirected to the login.

waylaidwanderer/node-chatgpt-api#192

When I entered "你好" in Chinese, an error message such as "Prompt empty or too short" appeared.
"你好" means hello, and its length is only 2 characters in Chinese.
When we use Chinese input, the length of the phrases we use can be very short, and in these cases, the program may indicate an error.

When I open the dev mode on my console , and click the submit bottom , then it report following issue
Im using the node version is v.19.0.0
Uncaught TypeError: crypto.randomUUID is not a function
at submitMessage (TextChat.jsx:155)
at handleKeyDown (TextChat.jsx:261)
at HTMLUnknownElement.callCallback (react-dom.development.js:4164)
at Object.invokeGuardedCallbackDev (react-dom.development.js:4213)
at invokeGuardedCallback (react-dom.development.js:4277)
at invokeGuardedCallbackAndCatchFirstError (react-dom.development.js:4291)
at executeDispatch (react-dom.development.js:9041)
at processDispatchQueueItemsInOrder (react-dom.development.js:9073)
at processDispatchQueue (react-dom.development.js:9086)
at dispatchEventsForPlugins (react-dom.development.js:9097)

It would be really nice to be able to choose between different models like gpt-3.5-tubo, text-davinci-003, (the upcoming) gpt-4....

It could be as an option in the customGPT

Thanks, the project is amazing!

When I ask a question using Bing, I find the following error:

ReferenceError: citationRegex is not defined
    at ask (/home/runner/chatGPT/chatgpt-clone/api/server/routes/askBing.js:99:50)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
    at async /home/runner/chatGPT/chatgpt-clone/api/server/routes/askBing.js:43:10

And there are no issues when I ask a question using Sydney.

Hi @danny-avila , I wanna if you need a PR containing a multi-user system? I have wrote a simple one and it's somehow extendable for future development.

For now, that's what in my implementation:

1. add a multi-user system, including login, logout and an API /api/me
2. UI will display username at the left bottom
3. Conversation will be connected to the username, so they will not see others' conversation.
4. The multi-user system is super simple, it based on session.user, containing {display, username}. the display will be used as display name, and the username will be used as a unique string in database.
5. The login and logout endpoint is not truly implemented, because I use OIDC to connect to my own user center server, so login and logout will redirect to the right place. (but I cannot provide the OIDC part. it's plugin design in order to using any other external auth system, just implement the login and logout, and set the session.user as expected.)

.

The three points on the upper left corner of the mobile endpoint did not respond. This is a small bug

have no idea on how to do that in server side, do you have any idea? @danny-avila

I think it would be better to set the client request url as a settable parameter since we need to replace all "http://localhost:3080" if the client and api are in different machine.

WSL2 Environment
ran docker instructions w/ updated tokens
set up a free instance of mongodb atlas e.g. connection string: mongodb+srv://user:[email protected]#s.mongodb.net/?retryWrites=true
upon finishing the first message in convo

happens for both bing and chatgpt models, exact same error message
"An error occurred. Please try again in a few moments.

Error message: Request failed with status code 429"
removed keys from chatgpt-clone-issue-1.txt

IMO, it's trying to save the response as a conversation but runs 429 "too many requests" while doing so

With waylaidwanderer/node-chatgpt-api#199, the customized system message is back; like custom GPT, we should support custom bing AI.

I tried changing model to gpt-4 but it gives me this error:

L'errore è questo:

Error message: Failed to send message. HTTP 404 - {
"error": {
"message": "This is a chat model and not supported in the v1/completions endpoint. Did you mean to use v1/chat/completions?",
"type": "invalid_request_error",
"param": "model",
"code": null
}
}

Is it possible to change the model? I'm facing some failure cases on that
@danny-avila

As title, I've been trying to implement @waylaidwanderers jailbreak client for Bing for a full day and my head is spinning lol.

All it needs is to pass jailbreakConversationId as true on first message, and track it and parentMessageId the same as you would conversationId usually, it doesn't require conversationId, invocationId or conversationSignature or anything like the original bing client does.

I can add in the jailbreakConversationId as true on the first message before it's sent, but I cannot seem to work out how it's building the convo object before the message is being sent.

Just looking for a bit of help on where to look so I can add it...

From Issue #15

Here is the copy from @Kristovich and his test:

Kristovich

I'm using the docker compose from da62e9a. I built the api and client containers as instructed on the readme.

I also added console.log(`Line 11, host${host}port${port}`); line inside api/server/index.js. When - HOST="0.0.0.0" is set I get:

api_1      |
api_1      | > [email protected] start
api_1      | > node server/index.js
api_1      |
api_1      | Line 11, host"0.0.0.0"port3080
api_1      | node:events:490
api_1      |       throw er; // Unhandled 'error' event
api_1      |       ^
api_1      |
api_1      | Error: getaddrinfo ENOTFOUND "0.0.0.0"
api_1      |     at GetAddrInfoReqWrap.onlookup [as oncomplete] (node:dns:107:26)
api_1      | Emitted 'error' event on Server instance at:
api_1      |     at GetAddrInfoReqWrap.doListen [as callback] (node:net:1934:12)
api_1      |     at GetAddrInfoReqWrap.onlookup [as oncomplete] (node:dns:107:17) {
api_1      |   errno: -3008,
api_1      |   code: 'ENOTFOUND',
api_1      |   syscall: 'getaddrinfo',
api_1      |   hostname: '"0.0.0.0"'
api_1      | }

If - HOST=0.0.0.0 then

api_1      | 
api_1      | > [email protected] start
api_1      | > node server/index.js
api_1      | 
api_1      | Line 11, host0.0.0.0port3080
api_1      | Server listening at http://0.0.0.0:3080

And while I'm at it: If - PROXY="" is present inside the docker-compose.yml, then the server will work but requests to openai or bing will fail with

api_1      | ask log {
api_1      |   model: 'chatgpt',
api_1      |   id: '[uuid]',
api_1      |   sender: 'User',
api_1      |   text: 'this is a test',
api_1      |   parentMessageId: undefined,
api_1      |   conversationId: undefined,
api_1      |   chatGptLabel: '',
api_1      |   promptPrefix: ''
api_1      | }
api_1      | TypeError [ERR_INVALID_URL]: Invalid URL
api_1      |     at new NodeError (node:internal/errors:399:5)
api_1      |     at new URL (node:internal/url:588:13)
api_1      |     at new ProxyAgent (/api/node_modules/undici/lib/proxy-agent.js:67:25)
api_1      |     at ChatGPTClient.getCompletion (file:///api/node_modules/@waylaidwanderer/chatgpt-api/src/ChatGPTClient.js:168:31)
api_1      |     at ChatGPTClient.sendMessage (file:///api/node_modules/@waylaidwanderer/chatgpt-api/src/ChatGPTClient.js:297:24)
api_1      |     at async askClient (/api/app/chatgpt-client.js:28:15)
api_1      |     at async /api/server/routes/ask.js:90:23 {
api_1      |   input: '""',
api_1      |   code: 'ERR_INVALID_URL'
api_1      | }

So I had to remove the PROXY env out of the docker-compose file for it to work.

I hope this helps.

First off. This is an amazing project. Thanks for putting this together.

I downloaded the latest update (this morning) and I get the following error when deploying with Docker Compose:

chatgpt-clone-api-1     | node:events:490
chatgpt-clone-api-1     |       throw er; // Unhandled 'error' event
chatgpt-clone-api-1     |       ^
chatgpt-clone-api-1     | 
chatgpt-clone-api-1     | Error: getaddrinfo EAI_AGAIN "localhost"
chatgpt-clone-api-1     |     at GetAddrInfoReqWrap.onlookup [as oncomplete] (node:dns:107:26)
chatgpt-clone-api-1     | Emitted 'error' event on Server instance at:
chatgpt-clone-api-1     |     at GetAddrInfoReqWrap.doListen [as callback] (node:net:1934:12)
chatgpt-clone-api-1     |     at GetAddrInfoReqWrap.onlookup [as oncomplete] (node:dns:107:17) {
chatgpt-clone-api-1     |   errno: -3001,
chatgpt-clone-api-1     |   code: 'EAI_AGAIN',
chatgpt-clone-api-1     |   syscall: 'getaddrinfo',
chatgpt-clone-api-1     |   hostname: '"localhost"'
chatgpt-clone-api-1     | }
chatgpt-clone-api-1     | 
chatgpt-clone-api-1     | Node.js v19.7.0

I have also tried filling in - HOST="" in the Docker Compose file with the proper hostname and it is still not working.

Any ideas?

There were some breaking changes with the latest node-chatgpt-api release. Will need time to review those changes on master, which may have suffered from auto-merging issues.

The now renamed and default main branch is confirmed to be working as expected so you can use this as the base.

Hey, I saw that you sponsored me! While checking out your profile, I came across your project and wanted to say that it looks great! I really like your prompt template feature.

Your idea of having an icon in the text input to switch between AI clients is brilliant, and I may borrow it for my own web client.

Thanks for the sponsor, and keep up the good work!

Presently, the prompt line does not have automatic focus by default. Consequently, you are required to manually click on the input field before they can begin typing, which leads to slower performance and disallows the use of keyboard-only interactions.

In contrast, in the original ChatGPT interface, as well as in the default Bing and Poe interfaces, the input field has automatic focus from the outset, allowing for immediate keyboard use without the need for any additional clicks.

bug 1:
When save and submit a new chatgptCustom, there is nothing added to the model menu, unless refresh

bug2:
when have multiple chatgptCustom, deleted one of them, it looks like another one is deleted. （but after refresh, everything got correct）

BTW, what's the differencee between save and submit? I got confused.

In testing this user script: https://greasyfork.org/en/scripts/461220-powerful-bing-chat-ai I found that Bing AI is less likely to refuse to answer when the chat is encoded in Base64 or ROT13. Supporting this on the client side, especially for "jailbroken" engines could provide an additional layer of opportunities for bypassing restrictions.

Adding Anthropic Claude bot to the list would be a great idea as it possesses unique strengths and weaknesses compared to ChatGPT or Sydney. Currently, it is only accessible through Quora Poe, but there exists a reverse-engineered API for it at https://github.com/vaibhavk97/Poe

api_1 | CLIENT RESPONSE {
api_1 | response: 'Hello there! Did you mean to type BAAI? If so, what can I assist you with regarding that term?',
api_1 | conversationId: '1f35b103-a084-4246-af3d-',
api_1 | parentMessageId: 'd275df1b-bae1-4128-97e7-',
api_1 | messageId: 'ae141143-74f6-4df3-b5a5-****',
api_1 | details: {}
api_1 | }
api_1 | TypeError: Cannot read properties of undefined (reading 'includes')
api_1 | at ask (/api/server/routes/ask.js:134:25)
api_1 | at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
api_1 | at async /api/server/routes/ask.js:51:10
api_1 | The client has disconnected.

Connected to MongoDB
npm ERR! path /api
npm ERR! command failed
npm ERR! signal SIGTERM
npm ERR! command sh -c node server/index.js

npm ERR! A complete log of this run can be found in:
npm ERR! /root/.npm/_logs/2023-03-09T08_22_49_452Z-debug-0.log

[email protected] start
node server/index.js

Server listening at http://localhost:3080
Connected to MongoDB

when new message is receiving, especially add new lines too quick, the auto scroll to end behaves to be hesitantly, and cannot scroll fast enough to the end.

this is mostly happened in mobile version. you can try write me a poem

here is a video. you can see the scroll cannot match the end starting from the second response, while responding.

I think it's because we call scrolltoend too much.
maybe we can detect scrolltop of the end element and skip some scrolltoend call if not changed. (might not fix completely)

I got this error when I using Customize ChatGPT Prompt

An error occurred. Please try again in a few moments.

Error message: Cannot read properties of undefined (reading 'includes')

When using custom name for a customGPT, the CSS style displayed on the web interface is incorrect.

i noticed when using official chatgpt, i will get title with the same language.
but in this project, only English title will be generated. even though i talk with chatgpt using Chinese