Dev-GPT is an experimental project developed by Jina AI. As an "automated development team," it aims to turn your ideas into reality through advanced artificial intelligence technology. Its core idea is: tell your AI team what microservice you want to build, and they will complete it for you. The team consists of a virtual product manager, developer, and DevOps, covering every step from concept to deployment.
The main capabilities include generating microservices, running microservices, deploying microservices, and deleting microservices. Generating a microservice requires providing a task description (optional), selecting a model (gpt-3.5-turbo or gpt-4, by default using the largest model you have permission for), and a local generation path (default ./microservice). The creation process takes about 5 to 15 minutes, during which GPT iteratively builds the microservice until it finds a strategy that allows the test scenarios to pass. In terms of cost, using GPT-4 costs about $0.50 to $3.00 per microservice, while using GPT-3.5-turbo costs about $0.05 to $0.30.
When running a microservice, it can be run locally in Docker; if Docker is not running, it will try to run without Docker. This command opens the playground in the browser for you to test the microservice. Deploying a microservice to the cloud requires a Jina account. Creating an account gives you some free credits that can be used for deployment ($0.025 per hour), and you can purchase more when the credits run out. Deleting a microservice can free up credits through a command.
In terms of technical implementation, Dev-GPT identifies multiple strategies for implementing the task and tests each strategy until it finds a viable one. For each strategy, it generates microservice.py (main implementation), test_microservice.py (test cases), requirements.txt (dependency list), and Dockerfile (container runtime and tests executed during build). It then tries to build the image, and if it fails, it uses the error information to fix it and retry. Once a successful strategy is found, it pushes the Docker image to the registry, deploys the microservice, and generates a Streamlit playground for testing. If it fails 10 times in a row, it moves on to the next approach.
In terms of usage, you need to set the OpenAI API key (requires access to gpt-3.5-turbo or gpt-4). If you want to enable microservice search of web content, you also need to set GOOGLE_API_KEY and GOOGLE_CSE_ID environment variables. If the OPENAI_API_KEY environment variable is set, the configuration step can be skipped; however, the key must have permission to use gpt-4, and the project is working to support gpt-3.5-turbo.
Applicable scenarios include generating various microservice examples, such as a compliment generator, extracting and summarizing news articles from URLs, chemical formula visualization, 2D rendering of 3D models, product recommendations, Hacker News search, an animal detector, a rhyme generator, a word cloud generator, table extraction, audio-to-mel-spectrogram conversion, a heatmap generator, a QR code generator, Mandelbrot set visualization, Markdown-to-HTML conversion, and more.
Limitations and notes: This project is an experimental version and may be unstable. An OpenAI key is required, and using gpt-4 works better but costs more. Deploying to Jina Cloud requires an account and credits. The generation process may get stuck and need to be interrupted and redone. At present, full support verification for Windows and Linux is not available, and multiple endpoints, stateful microservices, other large language models, etc. are not supported (these are on the to-do list). In addition, the player may print repeatedly, and there may be issues such as duplicate code in the code.
