ChatGPT Clone

The challenge

Putting an OpenAI key in a frontend isn't an option, and every model capability answers differently: JSON, text, a stream, audio or an image. The goal was an architecture where adding a new use case means adding one file, not rewiring half the app.

Implementation

• The backend is the only thing holding the API key

  NestJS acts as a proxy: the browser never talks to OpenAI directly. Each capability lives isolated in its own use case, so adding one doesn't touch the rest.

• Token-by-token streaming

  Long answers are relayed to the client as they arrive through a custom pipe over the OpenAI stream, which also shuts down cleanly if the client aborts mid-generation.

• Multimodal, not just text

  Spelling, translation and pros/cons in text; text to speech and audio transcription with uploads validated by type and size; plus image generation and variation.

• Assistant with persistent threads

  Alongside completions-based chat, a separate module uses the Assistants API, creating threads that keep conversation context across requests.

• Swappable provider via Ollama

  The same use cases have an alternative Ollama implementation for running models locally, and a public endpoint reports which providers a given deployment has available.

• History persisted in PostgreSQL

  Conversations and messages are stored with TypeORM, alongside JWT authentication, email password recovery and the app's own user accounts.

• Hardened API

  Rate limiting with Throttler, security headers via Helmet, compression, environment validation with Joi at boot and DTOs validated with class-validator.