Assistant AIs are the future and OpenAI revolutionized this movement with ChatGPT. While numerous methods exist to integrate them, this app commemorates the original styling of ChatGPT, with the ability to integrate any current/future AI models, while improving upon original client features, such as conversation search and prompt templates (currently WIP).
This project was started early in Feb '23, anticipating the release of the official ChatGPT API from OpenAI, and now uses it. Through this clone, you can avoid ChatGPT Plus in favor of free or pay-per-call APIs. I will soon deploy a demo of this app. Feel free to contribute, clone, or fork. Currently dockerized.
[Latest release (v0.0.4)](https://github.com/danny-avila/chatgpt-clone/releases/tag/v0.0.4) includes Resubmitting messages & Branching messages, which mirrors official ChatGPT feature of editing sent message, that then branches the conversation into separate message paths (works only with ChatGPT)
Full details and [example here](https://github.com/danny-avila/chatgpt-clone/releases/tag/v0.0.4). Message search is on the docket
Really thankful for all the issues reported and contributions made, the project's features and improvements have accelerated as result. Honorable mention is [wtlyu](https://github.com/wtlyu) for contributing a lot of mindful code, namely hostname configuration and mobile styling. I will upload images on next release for faster docker setup, and starting updating them simultaneously with this repo.
Many improvements across the board, the biggest is being able to start conversations simultaneously (again thanks to [wtlyu](https://github.com/wtlyu) for bringing it to my attention), as you can switch conversations or start a new chat without any response streaming from a prior one, as the backend will still process/save client responses. Just watch out for any rate limiting from OpenAI/Microsoft if this is done excessively.
Adds Sydney (jailbroken Bing AI) to the model menu. Thank you [DavesDevFails](https://github.com/DavesDevFails) for bringing it to my attention in this [issue](https://github.com/danny-avila/chatgpt-clone/issues/13). Bing/Sydney now correctly cite links, more styling to come. Fix some overlooked bugs, and model menu doesn't close upon deleting a customGpt.
I've re-enabled the ChatGPT browser client (free version) since it might be working for most people, it no longer works for me. Sydney is the best free route anyway.
Due to increased interest in the repo, I've dockerized the app as of this update for quick setup! See setup instructions below. I realize this still takes some time with installing docker dependencies, so it's on the roadmap to have a deployed demo. Besides this, I've made major improvements for a lot of the existing features across the board, mainly UI/UX.
Custom prompt prefixing and labeling is now supported through the official API. This nets some interesting results when you need ChatGPT for specific uses or entertainment. Select 'CustomGPT' in the model menu to configure this, and you can choose to save the configuration or reference it by conversation. Model selection will change by conversation.
Official ChatGPT API is out! Removed davinci since the official API is extremely fast and 10x less expensive. Since user labeling and prompt prefixing is officially supported, I will add a View feature so you can set this within chat, which gives the UI an added use case. I've kept the BrowserClient, since it's free to use like the official site.
The Messages UI correctly mirrors code syntax highlighting. The exact replication of the cursor is not 1-to-1 yet, but pretty close. Later on in the project, I'll implement tests for code edge cases and explore the possibility of running code in-browser. Right now, unknown code defaults to javascript, but will detect language as close as possible.
BingAI is integrated (although sadly limited by Microsoft with the 5 msg/convo limit, 50 msgs/day). I will need to handle the case when Bing refuses to give more answers on top of the other styling features I have in mind. Official ChatGPT use is back with the new BrowserClient. Brainstorming how to handle the UI when the Ai model changes, since conversations can't be persisted between them (or perhaps build a way to achieve this at some level).
Official ChatGPT use is no longer possible though I recently used it with waylaidwanderer's [reverse proxy method](https://github.com/waylaidwanderer/node-chatgpt-api/blob/main/README.md#using-a-reverse-proxy), and before that, through leaked models he also discovered.
Currently, this project is only functional with the `text-davinci-003` model.
> This is a work in progress. I'm building this in public. FYI there is still a lot of tech debt to cleanup. You can follow the progress here or on my [Linkedin](https://www.linkedin.com/in/danny-avila).
If your server cannot connect to the chatGPT API server by some reason, (eg in China). You can set a environment variable `PROXY`. This will be transmitted to `node-chatgpt-api` interface.
**Warning:** `PROXY` is not `reverseProxyUrl` in `node-chatgpt-api`
<details>
<summary><strong>Set up proxy in local environment </strong></summary>
Here is two ways to set proxy.
- Option 1: system level environment
`export PROXY="http://127.0.0.1:7890"`
- Option 2: set in .env file
`PROXY="http://127.0.0.1:7890"`
**Change `http://127.0.0.1:7890` to your proxy server**
</details>
<details>
<summary><strong>Set up proxy in docker environment </strong></summary>
set in docker-compose.yml file, under services - api - environment
```
api:
...
environment:
...
- "PROXY=http://127.0.0.1:7890"
# add this line ↑
```
**Change `http://127.0.0.1:7890` to your proxy server**
This project was originally created as a Minimum Viable Product (or MVP) for the [@HackReactor](https://github.com/hackreactor/) Bootcamp. It was built with OpenAI response streaming and most of the UI completed in under 20 hours. During the end of that time, I had most of the UI and basic functionality done. This was created without using any boilerplates or templates, including create-react-app and other toolchains. I didn't follow any 'un-official chatgpt' video tutorials, and simply referenced the official site for the UI. The purpose of the exercise was to learn setting up a full stack project from scratch. Please feel free to give feedback, suggestions, or fork the project for your own use.
This means my implementation or the underlying model may not behave exactly the same in some ways:
- Conversations are not tied to any user IDs, so if that's important to you, you should implement your own user ID system.
- ChatGPT's model parameters (temperature, frequency penalty, etc.) are unknown, so I set some defaults that I thought would be reasonable.
- Conversations are limited to roughly the last 3000 tokens, so earlier messages may be forgotten during longer conversations.
- This works in a similar way to ChatGPT, except I'm pretty sure they have some additional way of retrieving context from earlier messages when needed (which can probably be achieved with embeddings, but I consider that out-of-scope for now).
