How weirenjie3/fastgpt works
Project Type & Purpose
• FastGPT is a full-stack, monorepo web application for knowledge-base Q&A and visual workflow orchestration built on large language models (LLMs).
• It provides:
• A Next.js/React frontend (ChakraUI)
• A Node/TypeScript backend API (monolithic service layer + NestJS sandbox service)
• Data storage in MongoDB (metadata, sessions, permissions) and PostgreSQL with vector extensions (pgvector or Milvus) for embeddings
• Auxiliary Python microservices (file→text conversion, rerankers, fetch APIs)
• Deployment via Docker, Helm charts, Kubernetes
File Structure Highlights
• /packages
• global: shared TS types & utilities
• service: core backend API modules (core, support, common libs)
• web: standalone React/Next.js frontend packages used by project/app
• plugins: Node worker for external integrations (DuckDuckGo, Feishu, Doc2X, etc.)
• /projects/app: Next.js application combining /packages/web & /packages/service via APIs
• /projects/sandbox: NestJS-based “sandbox” microservice for secure code execution
• /python: several Python microservices for specialized tasks (fetching URLs, Office → text, BGE rerankers)
• /files/docker & /files/helm: infrastructure as code—Docker-compose definitions and Helm charts for Kubernetes
README Insights
• Tech stack called out: NextJs + TypeScript + ChakraUI + Mongo + Postgres (Vector plugin)
• Emphasis on LLM model orchestration, knowledge store ingestion, workflow (“Flow”) builder, OpenAPI endpoints, embed-share, voice I/O.
a. Main Components
• Frontend (Next.js app in /projects/app)
• API Gateway / Backend (packages/service HTTP controllers & business logic)
• Sandbox Microservice (NestJS in /projects/sandbox)
• Python Microservices (python/api, python/bge-rerank)
• Plugins Worker (packages/plugins runtime/worker.ts)
• Databases:
• MongoDB for metadata, user, session, logs, etc.
• PostgreSQL + pgvector (or Milvus) for vector store/embeddings
• External services:
• LLM providers via one-API (OpenAI, Baichuan, ChatGLM, etc.)
• Third-party data sources (URLs, Office files, search engines)
• DevOps: Docker, Docker-Compose, Helm, Kubernetes
b. Relationships & Data Flows
• Frontend ⟷ Backend API: REST/GraphQL calls
• Backend ⟷ Sandbox Service: RPC/HTTP to isolate code execution
• Backend ⟷ Python Services: HTTP calls (file conversion, rerank)
• Backend ⟷ Databases:
– Mongo client for system data
– Vector client for embeddings storage & retrieval
• Backend ⟷ External LLMs: via one-API abstraction layer
• Plugins Worker: subscribes to job queue or invoked by backend to run tool templates
c. Architectural Patterns & Principles
• Monorepo with package isolation, code-sharing via Lerna/pnpm workspace
• Layered architecture: API/controllers → core services → data stores → external adapters
• Microservices for sandboxed code execution and specialized algorithms
• Plugin architecture: dynamic tool/third-party integrations loaded at runtime
• Infrastructure as code: reproducible Docker/Helm deployments
d. Key Technologies
• Next.js, React, ChakraUI for UI
• Node.js + TypeScript for API (Express-style controllers under packages/service)
• NestJS for sandbox microservice
• Python (Flask/FastAPI) for auxiliary services
• MongoDB, PostgreSQL + pgvector, Milvus
• Helm charts, Docker-Compose, Kubernetes