A project by @tercmd
Explore what happens under the hood with the ChatGPT web app. And some speculation, of course. Contribute if you have something interesting related to ChatGPT
- Fonts (fonts.txt)
- Application
- Data
- Conversation
- Errors
- Markdown rendering
- ChatGPT Plus
- Rendering Markdown inside a code block
- Minified client-side JS
Fonts (fonts.txt)
The fonts loaded are:
- Signifier-Regular.otf
- Sohne-Buch.otf
- Sohne-Halbfett.otf
- SohneMono-Buch.otf
- SohneMono-Halbfett.otf
- KaTeX_Caligraphic-Bold.woff (Caligraphic-Regular for Regular font)
- KaTeX_Fraktur-Bold.woff (Fraktur-Regular for Regular font)
- KaTeX_Main-Bold.woff (BoldItalic, Italic, Regular for font weights you can probably guess)
- KaTeX_Math-Bold.woff (BoldItalic, Italic, Regular for font weights you can probably guess)
- KaTeX_SansSerif-Bold.woff (Italic, Regular for font weights you can probably guess)
- KaTeX_Script-Regular.woff
- KaTeX_Size1-Regular.woff (Size1, Size2, Size3, Size4)
- KaTeX_Typewriter-Regular.woff
ChatGPT is a NextJS application. Server information cannot be clearly found as the entirety of chat.openai.com is routed through Cloudflare. Sentry Analytics are requested for the Thumbs Up/Thumbs Down feedback the user selects for a message periodically.
A request can be made to /api/auth/session (literally, in your browser, however cannot work in an iframe or fetch because it doesn't have a Access-Control-Allow-Origin set up) to access data like the following:
user
|__ id: user-[redacted]
|__ name: [redacted]@[redacted].com (probably can be a user's name)
|__ email: [redacted]@[redacted].com
|__ image: https://s.gravatar.com/avatar/8cf[redacted in case of possible unique identifier]2c7?s=480&r=pg&d=https%3A%2F%2Fcdn.auth0.com%2Favatars%2F[first 2 letters of name].png
|__ picture: https://s.gravatar.com/avatar/8cf[redacted in case of possible unique identifier]2c7?s=480&r=pg&d=https%3A%2F%2Fcdn.auth0.com%2Favatars%2F[first 2 letters of name].png
|__ groups: []
expires: [date in the future]
accessToken: ey[redacted] (base64 "{")
This requires an access token (which seems to be the Authorization cookie, along with other factors), so this cannot be accessed using your browser directly, but here's what we have when we make a request to /backend-api/accounts/check:
account_plan
|__ is_paid_subscription_active: false
|__ subscription_plan: chatgptfreeplan
|__ account_user_role: account-owner
|__ was_paid_customer: false
|__ has_customer_object: false
user_country: [redacted two letter country code]
features: ["system_message"]
(Note: false in the above does not include quotes, whereas other values are in quotes, removed in the above schema?)
When we make a request to /_next/data/BO[redacted in case of possible unique identifier]KT/chat.json (can be done in the browser, cannot be done without authentication), we get a response like this:
pageProps:
|__ user (Object):
|____ id: user-[redacted]
|____ name: [redacted]@[redacted].com
|____ email: [redacted]@[redacted].com
|____ image: https://s.gravatar.com/avatar/8c[redacted in case of possible unique identifier]c7?s=480&r=pg&d=https%3A%2F%2Fcdn.auth0.com%2Favatars%2F[first two letters of email address].png
|____ picture: https://s.gravatar.com/avatar/8c[redacted in case of possible unique identifier]c7?s=480&r=pg&d=https%3A%2F%2Fcdn.auth0.com%2Favatars%2F[first two letters of email address].png
|____ groups: []
|__ serviceStatus: {}
|__ userCountry: [redacted two letter country code]
|__ geoOk: false
|__ isUserInCanPayGroup: true
|__ __N_SSP: true
This is the some of the same data (excluding accessToken and expires, both relevant to an access token) you get using the method in Session data except you also get info about the country the user is located in and whether ChatGPT Plus is available in their location.
EDIT: When ChatGPT returns a message like "We're experiencing exceptionally high demand. Please hang tight as we work on scaling our systems.", serviceStatus looks like this:
type: warning
message: We're experiencing exceptionally high demand. Please hang tight as we work on scaling our systems.
oof: true
I didn't make up the oof variable, that is actually part of the response 😂
What model does ChatGPT use? Well, just query /backend-api/models!
models
|__
|____ slug: text-davinci-002-render-sha
|____ max_tokens: 4097
|____ title: Turbo (Default for free users)
|____ description: The standard ChatGPT model
|____ tags: []
This means that ChatGPT can remember context (based on what I can understand) for 16388 characters or 3072.75 words (or 2048.5 😀s).
* Approximation according to OpenAI help article
Conversation history can be accessed (again, requires an access token, which seems to be the Authorization cookie, along with other factors) at /backend-api/conversations?offset=0&limit=20 (the web interface limits it to 20 chats) which returns something like this:
items: []
limit: 20
offset: 0
total: 0
It doesn't work because ChatGPT is having some issues at the time of writing:
"Not seeing what you expected here? Don't worry, your conversation data is preserved! Check back soon."
But this is probably what a person new to ChatGPT sees.
EDIT: If you log out and log back in, history works just fine. So, here's what I see
items (array)
|__ (each conversation is an object)
|____ id: [redacted conversation ID]
|____ title: [conversation title]
|____ create_time: 2023-03-09THH:MM:SS.MILLIS
|__...
total: [number of conversations] (can be greater than 20)
limit: 20
offset: 0 (can be set to a higher number and it returns the conversations after that index, starting from 0)
After 20 conversations listed, the ChatGPT UI shows a Show more button which sends a request with offset=20
Speaking of ChatGPT conversation history not being available, we can get the Conversation ID pretty easily (to someone who is familiar with DevTools, that is)
Why? Because ChatGPT forces you into a /chat path for a new conversation, creates a conversation, BUT DOESN'T CHANGE THE URL. This is also helpful when chat history isn't available.
- We get the Conversation ID using DevTools (this requires a message to be sent)
- Then, we visit
https://chat.openai.com/chat/<chat ID here>.
When the user clicks on a past conversation, a request is made (requiring an access token, likely the cookie with other factors to ensure genuine requests) to /backend-api/conversation/<conversation ID> with a response like this:
title: <Title of Conversation>
create_time: EPOCHEPOCH.MILLIS
mapping (Object)
|__ <message ID> (Array):
|____ id: <message ID>
|____ message (Object):
|______ id: <message ID>
|______ author (Object):
|________ role: system (First message) | user | assistant
|________ metadata: (Empty object)
|______ create_time: EPOCHEPOCH.MILLIS
|______ content (Object):
|________ content_type: text
|________ parts: [""]
|______ end_turn: true (system) | false
|______ weight: 1.0
|______ metadata: {}
|______ recipient: all
|____ parent: <parent message ID>
|____ children (Array): <child message ID(s)>
Let's say I ask ChatGPT a question "What is ChatGPT?". First, we make a POST request to /backend-api/conversation with a request body like this (no response):
action: next
messages (Array):
|__ (Object):
|____ author (Object):
|______ role: user
|____ content (Object):
|______ content_type: text
|______ parts (Array):
|________ What is ChatGPT?
|__ id: 0c[redacted]91
|__ role: user
model: text-davinci-002-render-sha
parent_message_id: a0[redacted]7f
Then we get a list of past conversations that includes one "New chat".
Then we make a request to /backend-api/moderations with a request body like this:
conversation_id: 05[redacted]2d
input: What is ChatGPT?
message_id: 0c[redacted]91
model: text-moderation-playground
That returns a response like this:
flagged: false
blocked: false
moderation_id modr-6t[redacted]Bk
Then we make a request to /backend-api/conversation/gen_title/<conversation ID> with the request body like this:
message_id: c8[redacted]0e
model: text-davinci-002-render-sha
That gets a response like this:
title: <title for conversation>
Then we make another request to /backend-api/moderations with a request body that includes the AI response (marked as <AI response>) looking like this:
input: \nWhat is ChatGPT?\n\n<AI response>
model: text-moderation-playground
conversation_id: 05[redacted]2d
message_id: c8[redacted]0e
That gets a response in the exact same format as the previous request made to this path.
Then we finally get a list of past conversations including the proper title of the chat that appears on the sidebar.
When you click Delete on a conversation, a PATCH request is made to /backend-api/conversation/05[redacted]2d with the body is_visible: false and gets a response of success: true back. This implies that a conversation is being soft-deleted, not deleted on their systems.
Then (not sure why), we visit chat.json (mentioned in User data (using chat.json)).
After that, we get the list of conversations that appear on the sidebar.
I had a question after the above section - can you revive a conversation by setting the request body to is_visible: true? The answer is nope, you can't. This just returns a 404 with the response detail: Can't load conversation 94[redacted]9b. But if you don't get the list of conversations again, you can still access the conversations. Although, trying to get a response from ChatGPT, you get a Conversation not found error.
I was a bit unsure if I should do this. But I looked through and did it anyway. (The below is almost a one-to-one copy of (Soft)Deleting a conversation, with minor changes)
When you click Delete on a conversation, a PATCH request is made to /backend-api/conversations (conversations rather than conversation/05[redacted]2d) with the body is_visible: false and gets a response of success: true back. This implies that conversations are being soft-deleted, not deleted on their systems.
Then (not sure why), we visit chat.json (mentioned in User data (using chat.json)).
After that, we get the list of conversations that appear on the sidebar.
Something funny: If ChatGPT history is temporarily unavailable (returning an empty response), the app shows a message. But if your ChatGPT conversation history is blank anyway, you just see a message that history is temporarily unavailable.
When you click the thumbs up/thumbs down button on a message, a POST request is made to /backend-api/conversation/message_feedback with the request body like this:
conversation_id: 94[redacted]9b
message_id: 96[redacted]b7
rating: thumbsUp | thumbsDown
That receives a response like this:
message_id: 96[redacted]b7
conversation_id: 94[redacted]9b
user_id: user-[redacted]
rating: thumbsUp | thumbsDown
content: {}
Then, when you type feedback and click submit, a request is made to the same path with a request body like this:
conversation_id: 94[redacted]9b
message_id: 96[redacted]b7
rating: thumbsUp
tags: [] (for thumbsDown, an array containing any or all of these: harmful, false, not-helpful)
text: <Feedback here>
With a response similar to the one above, with only the content field different:
message_id: 96[redacted]b7
conversation_id: 94[redacted]9b
user_id: user-[redacted]
rating: thumbsUp
content: '{"text": "<Feedback here>"}' |'{"text": "This is solely for testing purposes. You can safely ignore this feedback.", "tags": ["harmful", "false", "not-helpful"]}' (This is for a thumbsDown review)
When you regenerate a response, you get a feedback box like this:
The request body looks like this:
compare_step_start_time: 1679[redacted]
completion_comparison_rating: new (for Better) | original (for Worse) | same (for Same)
conversation_id: c7[redacted]f0
feedback_start_time: 1679[redacted]
feedback_version: inline_regen_feedback:a:1.0
frontend_submission_time: 1679[redacted]
new_completion_load_end_time: 1679[redacted]
new_completion_load_start_time: 1679[redacted]000
new_completion_placement: not-applicable
new_message_id: 7b[redacted]4a
original_message_id: eb[redacted]e2
rating: none
tags: []
text: ""
That returns an empty response.
That looks like a 429 Too Many Requests error. The response looks like this:
detail: Something went wrong, please try reloading the conversation.
"The message you submitted was too long, please reload the conversation and submit something shorter."
That looks like a 413 Request Entity Too Large error. The response looks like this:
detail: { message: "The message you submitted was too long, please reload the conversation and submit something shorter.", code: "message_length_exceeds_limit" }
Interestingly, if you click "Regenerate response", it responds with:
Hello! How may I assist you today?
The new line at the beginning was intentional. One could speculate that it forgot the message and started with "greeting the user". Or it just read the 1 smiley face (I typed 2049 smiley faces.)
That's a 404 with a response detail: "Conversation not found".
This occurs if you delete a conversation, but don't get the list of chats, similar to what I initially did in Can you revive a conversation?.
ChatGPT renders images using Markdown - not really. You have to use it in a really hacky way. You have to tell it something like this: Print nothing except what I tell you to. Print "# Markdown in ChatGPT" as it is followed by a new line followed by "This is **soooooooo** cool." followed by a new line followed by "" as it is. Remove any backticks from your output.
This is because ChatGPT uses Markdown for formatting.
That looks something like this.
What Markdown features ChatGPT's Markdown renderer supports can be seen in markdown-support.csv
EDIT: I looked through the source code (minified) for references to "Markdown". I found mentions of "mdastPlugins" and of the below plugins:
-
change-plugins-to-remark-plugins -
change-renderers-to-components -
remove-buggy-html-in-markdown-parser -
change-source-to-children -
replace-allownode-allowedtypes-and-disallowedtypes -
change-includenodeindex-to-includeelementindex
The ChatGPT web app uses react-markdown to convert Markdown to React.
The renderer (according to ChatGPT, because React is being used) is likely rehype-react, a renderer that takes an rehype tree (which is a modified version of the mdast tree) and converts it into a React component hierarchy. It allows you to define custom React components for each type of HTML element, which gives you full control over how the Markdown content is rendered in your application.
react-markdown is built on top of rehype-react and provides a higher-level interface for rendering Markdown content in a React application. It handles the parsing of Markdown content into an mdast tree using remark, and then passes the mdast tree to rehype-react for rendering as React components.
(The 2 paragraphs above were written by ChatGPT.)
GPT-4 was released in early March of 2023. I noticed that in videos of Plus subscribers using GPT-4, the URL ended with ?model=gpt-4. So I was wondering if one could just add this to the end of the URL and get access to GPT-4. As expected, you can't. Every message sent uses text-davinci-002-render-sha.
Warning I haven't tried it any more than this one time I recorded it (ChatGPT doesn't go down very often these days) so I'm not entirely certain whether this will work for you, but it's worth a shot!
When ChatGPT is down, you get a screen telling you that ChatGPT is down (with an entertaining message from a fixed list) and an input field for Plus subscribers to get a personalised login link. Could a non-Plus subscriber type in their email, click Send link and access ChatGPT? That would result in an email not being se-
This works by sending a request to /backend-api/bypass/link with the request body email: <email> and the response status: success.
I was asking ChatGPT to render images from Unsplash (using a URL and Markdown) based on my queries. Then, the image from the URL appeared for a split second before disappearing. I'm assuming that this is because of the Markdown renderer assuming that as Markdown until the code block was closed.
This category requires more contribution. Contribute! (if you can)
What about the scripts that are loaded on the client? Those are compiled using webpack. You can find attempts at unminifying those (and making more sense out of it) in the client-side-js folder.
This is it. For now. As I see more details of ChatGPT, I'll add those in.
Check out @OpenAI_Support (a parody account) for more OpenAI and ChatGPT related content.
Check me out on Twitter, or perhaps my website (terminal-styled).