====== Dify ======

**Dify** is an open-source agentic workflow platform that enables developers and non-technical users to build, deploy, and manage LLM-powered applications through a visual workflow designer and programmatic APIs. With over **134,000 GitHub stars**, Dify has become one of the most popular platforms for orchestrating AI agents and RAG pipelines.

| **Repository** | [[https://github.com/langgenius/dify|github.com/langgenius/dify]] |
| **License** | Apache 2.0 |
| **Language** | Python, TypeScript |
| **Stars** | 134K+ |
| **Category** | Agentic Workflow Platform |

===== Key Features =====

  * **Visual Workflow Studio** — Drag-and-drop interface for designing AI workflows, training agents, and configuring RAG systems
  * **Multi-Model Support** — Access, switch, and compare performance across dozens of LLM providers including OpenAI, Anthropic, open-source models
  * **RAG Pipeline** — Built-in retrieval-augmented generation engine that extracts, transforms, and indexes data from various sources into vector databases
  * **Agent Node System** — Autonomous decision-making nodes within workflows using ReAct, Function Calling, Chain-of-Thought, Tree-of-Thought, and custom strategies
  * **Prompt IDE** — Dedicated prompt orchestration interface for configuring and managing prompts
  * **MCP Integration** — Native Model Context Protocol support for accessing external APIs, databases, and services
  * **Backend-as-a-Service** — One-click deployment as APIs, chatbots, or internal business tools

===== Architecture =====

Dify employs a **Beehive modular architecture** where each component can be developed, tested, and deployed independently. The platform comprises three core operational layers:

  * **LLM Orchestration Layer** — Manages connections and switching between multiple large language models
  * **Visual Studio Layer** — Drag-and-drop interface for designing workflows and configuring agents
  * **Deployment Hub** — Enables publishing as APIs, chatbots, or internal tools

Model suppliers and models are configured declaratively using **YAML-based DSL**, standardizing the process of adding new models while maintaining API consistency across integration points.

<mermaid>
graph TB
    subgraph Client["Client Layer"]
        WebUI[Web UI]
        API[REST API]
        MCP[MCP Server]
    end
    subgraph Core["Core Engine"]
        WF[Workflow Engine]
        Agent[Agent Node]
        Prompt[Prompt IDE]
    end
    subgraph Models["LLM Orchestration"]
        OpenAI[OpenAI]
        Claude[Anthropic]
        OSS[Open Source Models]
    end
    subgraph Data["Data Layer"]
        RAG[RAG Pipeline]
        VDB[(Vector DB)]
        KB[Knowledge Base]
    end
    Client --> Core
    Core --> Models
    Core --> Data
    RAG --> VDB
    RAG --> KB
    Agent --> WF
</mermaid>

===== Agent Strategies =====

The Agent Node functions as a decision center within workflows, supporting multiple reasoning strategies:

  * **ReAct** — Chain of Think-Act-Observe cycles
  * **Function Calling** — Precise function-based tool invocation
  * **Chain-of-Thought (CoT)** — Step-by-step reasoning
  * **Tree-of-Thought (ToT)** — Branching exploration of reasoning paths
  * **Graph-of-Thought (GoT)** — Graph-based reasoning structures

Developers can create custom strategy plugins using CLI tools and customize configuration forms.

===== Code Example =====

<code python>
import requests

DIFY_API_KEY = "app-your-api-key"
BASE_URL = "https://api.dify.ai/v1"

def run_workflow(inputs, user_id="default"):
    response = requests.post(
        f"{BASE_URL}/workflows/run",
        headers={
            "Authorization": f"Bearer {DIFY_API_KEY}",
            "Content-Type": "application/json"
        },
        json={
            "inputs": inputs,
            "response_mode": "blocking",
            "user": user_id
        }
    )
    return response.json()

result = run_workflow({"query": "Summarize this document"})
print(result["data"]["outputs"])
</code>

===== References =====

  * [[https://github.com/langgenius/dify|Dify GitHub Repository]]
  * [[https://dify.ai|Dify Official Website]]
  * [[https://docs.dify.ai|Dify Documentation]]
  * [[https://dify.ai/blog/dify-rolls-out-new-architecture|Dify Architecture Blog Post]]

===== See Also =====

  * [[ragflow|RAGFlow]] — RAG engine with deep document understanding
  * [[langfuse|Langfuse]] — LLM observability and tracing
  * [[mem0|Mem0]] — Memory layer for AI agents
  * [[mcp_servers|MCP Servers]] — Model Context Protocol implementations