lywnl/ai-app-generation

This is a Java project that takes a natural-language description of an app and turns it into a deployable web application. The README compares it to Meituan NoCode, Bolt.new, and Lovable. The backend runs on Spring Boot 3.5.4 with Java 25, and the AI orchestration uses LangChain4j with LangGraph4j for agent workflows. The frontend is a separate Vue 3 project using Vite, Ant Design Vue, Pinia, and TypeScript. The platform can produce three forms of code, picked automatically by a Qwen-Turbo router model: a single HTML file, a multi-file static project, or a full Vue engineering project. When the chosen mode is the Vue project, an agent calls six file tools to lay out and edit the project: FileWrite, Read, Modify, Delete, DirRead, and Exit. A parallel image-collection agent runs four tools at once for cover art, logo, illustrations, and diagrams, pulling from Pexels, Alibaba's wan2.2 logo generator, Mermaid, and unDraw. Retrieval-augmented generation is built in. Prompt context is pulled from a PgVector store, embedded with DashScope text-embedding-v4, and re-ranked with gte-rerank-v2 before being attached. The main code generation calls go to DeepSeek-Chat and stream back over Server-Sent Events using a Reactor Flux. Generated artifacts are hosted as static files behind Nginx so the result is reachable as a URL, and the project can also be downloaded as a ZIP. Several supporting systems are documented. A two-layer cache pairs Caffeine in process with Redis. Prompt safety runs as an @PromptSafetyCheck annotation plus an AOP aspect plus LangChain4j's InputGuardrail. User sessions persist in Redis for thirty days via Spring Session, with an @AuthCheck aspect for role gating. Observability covers Actuator metrics scraped by Prometheus with a Grafana dashboard, and webpage screenshots run via Selenium and Chromium and upload covers to Tencent COS. Setup requires JDK 25, Maven, Node 20, MySQL 8, Redis 7, and PostgreSQL 16 with the pgvector extension. The README lists the API keys needed before start: DeepSeek, DashScope, Pexels, and Tencent COS. The backend defaults to port 9025 and the frontend dev server to 5173, with Knife4j serving OpenAPI docs.