# Design ChatGPT — System Design Mock Interview

Aced (formerly Exponent) publishes a public search API over its indexed pages; the guide is at https://www.aced.io/agents/agent-access.

Source: https://www.aced.io/blog/design-chatgpt-system-design-mock-interview
Source observed: 2026-11-05T00:17:04.231Z
Last checked: 2026-11-05T00:17:04.231+00:00
Indexed: 2026-11-06T23:07:13.623+00:00

---

# Design ChatGPT — System Design Mock Interview

[System Design](https://www.aced.io/blog/tag/system-design)

By Aced Team · Last updated a month ago

We use this mock interview to walk through the design of ChatGPT, a conversational system. The design covers conversation lifecycle operations, message ratings, input sanitization, model serving, storage, scalability, security, and model improvement.

## Scope and functional requirements

We consider these core capabilities:

- Create, update, view, and delete conversations.
- Send messages within a conversation.
- Rate individual responses with a thumbs-up or thumbs-down.
- Use ratings as feedback for model training and retraining.
- Accept text-based inputs in English.
- Sanitize inputs to remove profanity and detect insults before processing.

## Non-Functional Requirements

Important non-functional requirements include latency, security, and scalability. Backend processing can make server response latency long. Login flows and rate-limiting procedures help prevent distributed denial-of-service attempts.

The rate limiter must scale so that we can serve several users simultaneously. We need a scalable database for conversation storage; we identify NoSQL as an ideal option for this design.

### Storage estimate

We estimate an average message size of **100 bytes**:

- **200 million messages per day**: approximately **20 GB per day**.
- Approximately **7.3 terabytes annually**.
- Approximately **76 terabytes over ten years**.

## High-Level Design

The central component is a conversation service that manages communication between users and the ChatGPT model. Before a message is processed, we sanitize and analyze it against the applicable standards.

The model's response is returned to the user and stored in a database as part of a conversation. A thumbs-down rating indicates that the model may need retraining. We also use a risk model to assess whether user ratings are legitimate.

## Conversation Service

We expose the conversation service as a REST API. Its operations include:

- Creating a conversation.
- Deleting a conversation.
- Viewing a conversation.
- Sending a message.

Each conversation has a unique ID, and each message has its own ID. Users can rate each message with a thumbs up or thumbs down.

We store the data in a NoSQL database. The conversation table contains conversation IDs. Each message contains:

- An ID
- Text
- Author
- Parent

## ChatGPT Model

We use a Transformer model to predict natural sequences of words. The model is trained on internet data, including websites, books, and Wikipedia links, to generate semantically meaningful and grammatically correct replies.

For token or word selection, the model can use approaches such as:

- Top K
- Greedy selection
- Nucleus temperature

These approaches represent different trade-offs in selecting a prediction.

## Chatbot Design

We initially train the model with a dataset of question-and-answer pairs. Since it is impossible to train on every possible question-and-answer combination, we use a reward model to select better responses.

The reward model is trained with reinforcement learning. The chatbot receives rewards for appropriate responses and penalties for inappropriate responses. The reward model considers the emotion and tone of the question and answer and continues training to improve accuracy.

The design is expected to support multiple input and output formats, including images, audio, and video. This enables natural-language responses even when the initial training dataset is not large.

## Other Considerations

- Challenges and limitations of ChatGPT technology, including potential bias.
- User testing and feedback to improve accuracy and functionality.
- Additional natural-language-processing techniques, such as sentiment analysis and named entity recognition.
- Trade-offs among model-training and selection approaches, including top K versus nucleus temperature.
- Ethical considerations for conversational agents, including privacy and possible effects on employment in certain industries.

## Conclusion

Designing ChatGPT requires both functional requirements—creating, updating, viewing, and deleting conversations—and non-functional requirements such as latency, security, and scalability. The design uses NoSQL storage and a Transformer model trained on internet data. Supervised fine-tuning, a reward model, and reinforcement learning improve the chatbot's natural-language responses, while the architecture is intended to support image, audio, and video formats.

## Continue learning with Aced

We provide free interview guides, insider tips, and courses through Aced, formerly Exponent. Our [complete system design interview course](https://www.aced.io/courses/system-design-interviews?ref=exponent-blog.ghost.io) is described as trusted by **19,000+** software engineers, engineering managers, technical program managers, and other technology professionals.

Watch the complete [Design ChatGPT](https://www.aced.io/courses/system-design-interviews/design-chatgpt?ref=exponent-blog.ghost.io) answer, or [create a free account](https://www.aced.io/signup?src=blog).

## Related articles

- [System Design Interview: Load Balancers](https://www.aced.io/blog/load-balancers-system-design-interview)
- [Design Food Delivery System like Uber Eats (Mock Interview)](https://www.aced.io/blog/design-food-delivery-uber-eats-system-design-interview)
- [System Design Interview Prep & Questions (2026 Guide)](https://www.aced.io/blog/system-design-interview-guide)
- [System Design Interview: Design a Rate Limiter](https://www.aced.io/blog/rate-limiter-system-design)

---

[Read this page’s summary](https://www.aced.io/agents/api/public/v1/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fblog%2Fdesign-chatgpt-system-design-mock-interview&trace=t_679ff9ab98f3)

## Ask a follow-up about this page

Need a more specific answer? Ask the exact question you have. It returns a fresh, cited answer from Aced (formerly Exponent)’s knowledge base. No account or API key required.

Open this URL with your own question, URL-encoded. Include your question in q. All research inputs are optional; no account or setup.

`GET https://www.aced.io/agents/agent/faq?q={question}&trace=t_679ff9ab98f3`

[Tool guide and examples](https://www.aced.io/agents/agent-access) · [OpenAPI](https://www.aced.io/agents/api/public/v1/openapi.json) · [Page TLDR form](https://www.aced.io/agents/agent-tools/page-tldr?url=https%3A%2F%2Fwww.aced.io%2Fblog%2Fdesign-chatgpt-system-design-mock-interview). Opening a form does not execute a tool.