langgenius/dify

Dify is an open-source platform for building production-grade AI applications through a visual drag-and-drop workflow builder. Instead of writing boilerplate code to chain LLM calls, manage prompts, and wire up retrieval pipelines, developers lay out their logic on a canvas -- connecting model nodes, tool calls, conditional branches, and human-in-the-loop checkpoints into executable workflows. The result is a system that can go from prototype to production deployment without rewriting the orchestration layer. The platform supports hundreds of LLM providers out of the box: OpenAI GPT models, Anthropic Claude, Mistral, Llama 3, Qwen, and any provider exposing an OpenAI-compatible API. Switching between models is a dropdown change, not a code refactor. This provider-agnostic design means teams can start with a cloud API, benchmark against alternatives, and migrate to self-hosted models without touching their workflow logic. Dify ships with a full RAG pipeline built in. Upload PDFs, presentations, or plain text, and the system handles chunking, embedding, vector storage, and retrieval. Version 1.12.0 introduced Summary Index, which attaches AI-generated summaries to document chunks so semantically related content clusters together during retrieval. Version 1.13.0 added multimodal retrieval that unifies text and images into a single semantic space for vision-enabled reasoning. The agent capabilities layer supports both function-calling and ReAct-style reasoning with over 50 built-in tools including Google Search, DALL-E, Stable Diffusion, and WolframAlpha. Since v1.0.0, all models and tools have been migrated to a plugin architecture, so extending Dify with custom integrations no longer requires forking the core codebase. On the operational side, Dify includes a Prompt IDE for comparing model outputs side-by-side, LLMOps-grade logging and performance monitoring, and a Backend-as-a-Service API layer that lets frontend applications consume workflows through REST endpoints. The v1.13.0 release added a Human Input node that pauses workflows for human review, enabling approval gates and content moderation loops directly within automated pipelines. Deployment is flexible: Docker Compose for quick self-hosted setups (minimum 2 CPU cores, 4 GB RAM), Kubernetes via five community-maintained Helm charts, Terraform templates for Azure and Google Cloud, and AWS CDK for infrastructure-as-code deployments. Dify Cloud offers a managed option with 200 free GPT-4 calls to get started. Enterprise customers including Volvo Cars and Kakaku.com run Dify in production -- Kakaku.com reported 75% employee adoption with nearly 950 internal AI applications built on the platform.