全栈 AI 平台Full-stack AI platform· 2026
LOOM-PLUS
可视化 AI 工作流编排引擎Visual AI workflow orchestration engine
把工作流画布、RAG 知识库、SSE 对话和失败补偿串成一套可运行的全栈系统。A full-stack AI workflow platform that connects a visual canvas, RAG knowledge base, streaming chat, and failure compensation.
案例摘要Case summary
目标用户Audience
想学习或验证 AI 工作流想法，但不想被重型平台和生态绑定的开发者、产品经理。Developers and product managers who want to learn or validate AI workflow ideas without being locked into heavyweight platforms and ecosystems.
我的角色Role
负责产品定义、功能范围、RAG 与工作流方案取舍和交付验收；全部代码由 AI 编程工具完成。Owned the product definition, feature scope, RAG and workflow tradeoffs, and acceptance decisions; all code was produced with AI coding tools.
交付结果Outcome
形成从画布编排、知识库检索到流式对话的完整链路，并保留真实截图和测试脚本。Delivered a complete chain from canvas orchestration and knowledge-base retrieval to streaming chat, backed by real screenshots and test scripts.
产品洞察Product insight
很多 AI 工作流平台要么绑定生态，要么部署成本过高。这个项目验证的是：个人开发者能否用更轻的方式理解、搭建并验证 AI 工作流产品的核心能力。Many AI workflow platforms are either locked into an ecosystem or too expensive to deploy. This project tests whether an individual developer can understand, build, and validate the core capabilities of an AI workflow product in a lighter way.
项目简介Overview
LOOM-PLUS 是一个面向 AI 应用构建的工作流平台，支持拖拽式 DAG 画布、7 类节点、知识库检索、混合排序和流式对话。它不是演示页，而是包含前端、后端、数据库、缓存、Docker 和验证脚本的完整项目。LOOM-PLUS is a workflow platform for AI application builders, with a drag-and-drop DAG canvas, 7 node types, knowledge-base retrieval, hybrid ranking, and streaming chat. It is not a demo page — it is a complete project with frontend, backend, database, cache, Docker, and verification scripts.
技术栈Tech stack
结果证据Evidence
仓库包含前端、后端、Docker、PostgreSQL、Redis 与 RAG 相关模块，可追溯完整实现。The repo includes frontend, backend, Docker, PostgreSQL, Redis, and RAG modules, so the full implementation is traceable.
docs/assets 中保留 workflow-canvas、hybrid-search、knowledge-base、chat-streaming 等真实界面截图。docs/assets keeps real UI screenshots of workflow-canvas, hybrid-search, knowledge-base, and chat-streaming.
scripts/test_health.js、test_rag.js、test_workflow_execution.js 覆盖健康检查、检索和工作流执行路径。scripts/test_health.js, test_rag.js, and test_workflow_execution.js cover health checks, retrieval, and workflow execution paths.
项目截图Screenshots
关键决策Key decisions
BM25 和向量相似度分数尺度不同，线性加权很难调准。BM25 and vector similarity scores use different scales, so linear weighting is hard to tune.
改用 RRF 只融合排名，降低不同检索通道分数不可比带来的噪声。Switch to RRF, which fuses rankings only, reducing noise from incomparable scores across retrieval channels.
工作流既要允许有限循环，又要防止死循环拖垮执行。A workflow needs to allow bounded loops while preventing infinite loops from dragging down execution.
用每节点访问次数上限和全局步数上限做双重保护。Use a per-node visit cap plus a global step cap as double protection.
想看完整的代码和实现细节？Want the full code and implementation details?
在 GitHub 上查看View on GitHub