> For the complete documentation index, see [llms.txt](https://dongzeli95s-organization.gitbook.io/swe-interview-handbook/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dongzeli95s-organization.gitbook.io/swe-interview-handbook/system-design/mocks/design-chatgpt.md).
# Design ChatGPT
## Functional Requirement:
1. Create conversation
2. Send message
3. Delete conversation
4. Language preference?
5. Do we support images/attachment?
6. Thumb up/down?
7. Login system? Rate Limiting?
8. How about context window? In the conversation?
## Non-functional Requirement:
1. Low Latency
2. Highly available
3. Scalable
## Scale
10M user/day -> 5 conversations, 4 messages -> 200M messages per day.
200M messages -> 100 bytes -> 20G per day -> 73TB (10 years)
## High Level Design
## GPT RLHF
### Fine tuning with RL
Proximal Policy Optimization
1. Policy is a language model that takes in a prompt and returns a sequence of text (just probability distributions over text)
## GPT RLHF
### Fine tuning with RL
Proximal Policy Optimization
1. Policy is a language model that takes in a prompt and returns a sequence of text (just probability distributions over text)
