How isec-ahu/llm4workflow works
Project type and purpose
• LLM4Workflow is a full-stack, LLM-driven web application that ingests a user’s natural-language workflow description, uses an LLM (OpenAI) and a vector store to generate and refine a workflow model, stores/retrieves API knowledge in Postgres, then emits an executable workflow in XML for deployment (e.g., on EdgeWorkflow).
File structure
• Top-level “backend” directory: Python LangChain service with
• server.py (FastAPI or LangChain serve entry point)
• db.py (Postgres interactions)
• vectorStore.py (embeddings storage/retrieval)
• create_game.py, rag.py, json_loader.py, write_xml.py, write_dag.py (pipeline stages)
• schema.py (Pydantic/OpenAPI schemas), utils.py
• init.sql, api.json (metadata)
• Docker-Compose for Postgres
• Top-level “frontend” directory: React + Vite + TypeScript app
• API layer under src/api
• Hooks under src/hooks
• Views/components under src/view
• Assets, styles, renderer utilities
• Tailwind config, ESLint, tsconfig, Vite config
README for additional insights
• Four main backend stages:
1) One-time API Knowledge Base Preparation
2) Task Extraction & Rewriting
3) API Retrieval
4) Workflow Model Generation (XML output)
• External dependencies/services:
• OpenAI API (LLM)
• Optional LangSmith tracing
• EdgeWorkflow (runtime evaluation platform)