Natural Language Generation of Automated Workflows: An Intelligent n8n Workflow Construction Platform Based on Large Language Models

This article introduces the open-source project n8n_workflow_generator, whose core is using generative AI to automatically convert natural language instructions into executable n8n workflows, solving the learning threshold problem for non-technical users using n8n. The project fine-tunes the Qwen2.5-Coder-7B model with QLoRA, trained on a dataset of over 4000 public n8n workflows, achieving an effective JSON generation rate of 93.3%. The platform supports natural language conversion, intelligent integration recognition, visual editing, and multi-method export, lowering the threshold for automation and accelerating development processes.

In digital work environments, n8n is a popular open-source workflow automation tool, but beginners and non-technical users face many thresholds: needing to understand node functions and configurations, handle authentication manually, master branch logic, and n8n's unique JSON structure. This project aims to use large language model capabilities to allow users to describe requirements in natural language and automatically generate corresponding n8n workflows, solving the above pain points.

1. Natural Language to Workflow: For example, if a user inputs "read Gmail emails daily, summarize with Gemini, save to Google Sheets, and send Telegram notifications", the system outputs a complete JSON containing Trigger, Gmail, Gemini, Google Sheets, and Telegram nodes.
2. Intelligent Integration Recognition: Automatically identifies services in requirements, supporting communication tools (Gmail, Slack, Telegram), data storage (Google Sheets, Airtable), AI services (OpenAI, Gemini), etc. Users can choose preferred services to optimize configurations.
3. Visualization and Export: Generated workflows can be dragged and edited on the canvas, connections adjusted, and support downloading JSON, copying to clipboard, importing into n8n, or deploying via API.

Dataset Construction: Collected 4033 unique workflows from repositories like Zie619/n8n-workflows, covering 495 node types, generating 32264 prompt-workflow paired data; preprocessing includes removing noise, standardizing structure, and filtering damaged files. Prompt Enhancement: Generate multiple variant prompts for each workflow (e.g., different descriptions of the same Gmail-Slack workflow) to improve generalization. Model Training: Selected the Qwen2.5-Coder-7B-Instruct base model, using QLoRA fine-tuning (Unsloth framework, context length 6144, LoRA rank 16, learning rate 2e-4). After 1 epoch of training, the validation loss was 0.3777, and the test set achieved an effective JSON generation rate of 93.3%.

System Flow: User prompt → Intent understanding → Integration detection and selection → Prompt refinement → Fine-tuned model → Workflow JSON generation → Validation → Visual rendering → Export. Tech Stack: Frontend (React, Vite, SVG canvas, Lucide Icons); Backend (Python, Hugging Face Transformers, Unsloth, Pandas/NumPy); Supports multiple LLM providers (OpenAI, Gemini, Groq, Ollama local models).

1. Lower Threshold: Non-technical users do not need to learn node configurations and can build automated processes with natural language.
2. Accelerate Development: Experienced n8n users can quickly generate workflow skeletons and then fine-tune them, improving efficiency.
3. Knowledge Reuse: Trained on large-scale datasets, the model learns best practices, ensuring the quality of generated code.

Future Plans: Multi-agent workflow planning, automatic credential mapping, workflow execution simulation, reinforcement learning optimization, multi-branch logic processing, automated compliance checks. Industry Significance: Demonstrates the potential of generative AI in workflow automation: natural language as code, domain-specific models, human-machine collaboration mode, evolution from low-code to conversational development. Provides a complete technical reference (data collection → preprocessing → training → visualization) for similar systems.