Dify

1. Introduction

1.1. Platform Positioning

Dify is an open-source LLM application development platform built by LangGenius, Inc. Its core proposition is: integrating agent workflows, RAG pipelines, model management, observability, and a plugin marketplace into a single collaborative workspace, letting teams go from prototype to production without rebuilding their tech stack.

Within the six-layer model and conceptual boundaries of the parameter card, Dify is an Agent Platform — Harness's productized wrapper: it layers UI, tenancy, billing, and publishing on top of L1~L3. This is the most fundamental difference between it and Framework-class solutions such as LangGraph, Claude Agent SDK, and ADK.

A memorable analogy is often cited: if WordPress democratized website building, Dify is doing the same for LLM application development — turning "writing code" into "dragging and dropping on a canvas."

1.2. Basic Information Card

1.3. Development Timeline

1.4. Position in the AI Harness System

Per the parameter-card boundary table, Dify clearly falls into the Agent Platform cell:

Key judgment: Dify is the only platform in this group with deliverable capabilities across all six layers L1~L6, at the cost of being shallower than specialized solutions at each layer.

2. Glossary

3. Feature Overview

3.1. Four Application Forms

3.2. Division of Labor Between Workflow and Chatflow

This is the pair of concepts most often misunderstood in Dify; the official explanation follows:

Selection rule of thumb: choose Chatflow for chatbots and Workflow for asynchronous batch processing.

3.3. Knowledge Base and RAG Pipeline

RAG is widely regarded as Dify's strongest capability; its complete pipeline is: document ingestion → chunking → vectorization → indexing → retrieval → reranking.

Comparison of three chunking strategies:

Parent-child chunking is Dify's differentiator: child chunks handle precise matching, while parent chunks provide the full context — for example, when a code snippet (child chunk) is hit, the complete function definition (parent chunk) is returned.

Supported input sources: PDF, TXT, Markdown, JSON, DOCX, XLSX, CSV, HTML; built-in web crawler (enter a URL); Notion / Confluence API sync; Knowledge API batch writes; OCR preprocessing for scanned PDFs.

Knowledge Pipeline further moves RAG's ETL onto the canvas: each step (parsing, cleaning, chunking, enrichment, vectorization) becomes a configurable node, with data sources covering Google Drive, Notion, Confluence, AWS S3, and Slack.

3.4. The Two Ways to Build an Agent

This recommendation is worth noting: Dify does not officially recommend building production multi-agent systems on "autonomous Agent Apps," but rather recommends demoting agent behavior to a single node within a workflow. This aligns with the core spirit of AI Harness — containing an uncertain core with a deterministic container.

3.5. Bidirectional MCP Capability

Since v1.6.0 (2025-07), Dify supports bidirectional MCP capability:

• As an MCP Client: agents inside Dify can directly call any external MCP server (official examples: Linear, Notion, Zapier), using HTTP-based MCP (protocol version 2025-03-26), replacing per-service integration code;
• As an MCP Server: any Dify application (Chatflow, Workflow, Agent) can be published as an MCP server, called by Claude Desktop, Cursor, or any MCP client.

The strategic implication of the latter is greater: Dify shifts from a "agent-building tool" to "a tool provider for agents in other ecosystems."

3.6. The Human Input Node

Introduced in v1.13.0, it is a key capability of Dify for regulated industries:

• A running workflow can pause to wait for human review;
• Reviewers respond via the Web app or email;
• Variables can be edited before the workflow continues;
• It can route to different branches via configurable buttons (approve / reject / escalate);
• A timeout is configurable (default 3 days); on timeout it takes the fallback branch.

Compared with LangGraph's interrupt, the Human Input node is more structured: pause points are defined at design time, rather than at arbitrary nodes at runtime. For scenarios such as legal review, compliance checks, and medical triage, this difference is decisive.

3.7. Plugin and Model Layer

• Since v1.0.0, models and tools are decoupled from the core as plugins, installed from the Marketplace;
• Supports hundreds of proprietary and open-source models: GPT, Claude, Gemini, Mistral, Llama, DeepSeek, Qwen, and more, as well as any OpenAI-compatible endpoint; self-hosted Ollama and Xinference models are first-class citizens;
• Models can be mixed within a single workflow: route cheap tasks to small models and complex reasoning to frontier models. This itself is a form of cost governance.

3.8. Observability and Publishing

• Observability: app logs, performance analysis, annotation feedback into datasets, latency and usage data; the debugger has better visualization of LLM token usage; can integrate Langfuse, Opik, Arize Phoenix.
• Publishing: each app can be published as a hosted Web app, an API endpoint, a web embedding component, or an MCP-compatible tool.

4. Platform Architecture

图 4-1｜Dify 自托管技术栈：从前端到异步 Worker 的分层架构

数据来源：基于本文分析绘制的示意图。

4.1 Deployment Forms

4.2 Self-Hosted Technology Stack

The official Docker Compose spins up a full set of services, with a minimum of 2 CPU cores + 4 GiB memory:

The async Worker architecture is the key differentiator: all time-consuming operations (PDF parsing, vector indexing, large file processing) run in background queues and do not block the API.

Production recommendations: replace with a dedicated vector-database cluster (Qdrant / Milvus), use external PostgreSQL, switch to S3-compatible storage, and use nginx + HTTPS.

4.3 Processing Flow of a Chatflow Request

1. The user message enters; session variables and memory are loaded;
2. If a knowledge retrieval node is present: query rewriting → vector / keyword / hybrid retrieval → reranking → context injection;
3. Routing to the corresponding model (different models may be selected per node within the workflow);
4. If an Agent Node is present: run the ReAct / Function Calling loop within that node, calling 50+ built-in tools or external MCP servers;
5. If a Human Input node is present: pause and wait for review (variables can be edited, branches or timeout fallback can be used);
6. Conditional branch / parallel / iteration nodes execute according to canvas semantics;
7. Streaming output to the Web app or API;
8. Logs and usage data are written for subsequent annotation and optimization.

5. Harness Design

5.1. Six-Layer Capability Overview

5.2. L1 Context Engineering Layer

Dify's L1 is the most fully productized layer within this group, because it turns retrieval engineering into visual configuration:

• Three chunking strategies are available, with parent-child chunking being the differentiator;
• Retrieval methods cover vector, keyword / BM25, and hybrid, with optional reranking;
• Knowledge Pipeline makes every ETL step visible and tunable;
• Per-model routing within workflows is itself part of context engineering — handing contexts of different complexity to models of different tiers.

Shortcomings:

• No context compaction primitive;
• No progressive disclosure (contrast Claude Agent SDK's Skills; Dify Agent Beta is just beginning to introduce Skills packaging);
• Sorting and trimming strategy after retrieval into context is still fairly fixed.

5.3. L2 Tools and Execution Layer

• Built-in tools: 50+, including Google Search, DALL·E, Stable Diffusion, WolframAlpha, and more;
• Plugin marketplace: model providers, tools, data sources, and MCP integrations installed uniformly; teams can reuse approved plugins (itself a governance action);
• Bidirectional MCP: can call external services and be called externally;
• Sandbox: code execution sandbox; the v1.16.0 Dify Agent (Beta) runs in a Linux sandbox.

A risk that must be highlighted: the Dify Agent Beta release notes explicitly warn that it should only be opened to trusted, non-malicious users. This is the inherent risk of a shell-based agent and a direct manifestation of the L2-layer principle "the more powerful the capability, the larger the attack surface."

5.4. L3 Orchestration and Control Layer

On the tension between flexibility ↔ predictability, Dify sits on the more-predictable side:

• The canvas is an explicit graph and control flow is visible — close to LangGraph in this respect;
• But it is less expressive than a code graph: complex conditional logic on a canvas easily degenerates into "spaghetti wiring";
• Iteration and parallel nodes are provided, but loop semantics are less rigorous than LangGraph's Pregel-style execution;
• The Human Input node fixes pause points at design time, making predictability higher than LangGraph's arbitrary-node interrupt, but with lower flexibility.

The official recommendation to "build production multi-agent systems with Agent Node rather than Agent App" is essentially a deliberate choice of the predictability end of this tension axis.

5.5. L4 Memory and State Layer

For comparison: LangGraph has Checkpoint + Thread + Time Travel; ADK has Session + Memory + Artifact (versioned). Dify only has "session-level" capability in this layer, with no run-level checkpoints or artifact versioning. This means long tasks cannot be resumed from the middle after a crash.

5.6. L5 Evaluation and Observability Layer

What exists:

• App logs and performance analysis;
• Annotation feeding back into improved datasets — a simple implementation of flowing production data back into the optimization loop;
• Latency and usage data;
• LLM token usage debugger (public reviews consider it superior to comparable tools);
• Can integrate third-party observability tools such as Langfuse, Opik, Arize Phoenix.

What is missing:

• No native Golden Dataset / regression-set mechanism (annotation partially substitutes, but is not the same thing);
• No automated workflow to "run the regression set and gate on thresholds with every release";
• No A/B and online metrics system.

Hence the "medium-strong" judgment — observability is sufficient, while evaluation requires external tools or manual annotation.

5.7. L6 Governance and Security Layer

There is also a license-level governance rule: the Dify Open Source License prohibits operating a multi-tenant hosted service without written permission and prohibits removing the front-end logos and copyright notices. This is not a technical guardrail, but it is the hardest business boundary of this platform — to commercialize multi-tenant SaaS on top of Dify's source code, you must first obtain a commercial license.

5.8. Concrete Manifestations of the Three Inherent Tensions

6. Real-World Cases

The following customer testimonials are all published publicly by Dify's official website; they are customer statements cited by the vendor, not independently verified, and none include quantifiable effect metrics.

Case 1: Volvo Cars (Head of AI & Data APAC)

"In this climate of perpetual beta, tools enabling rapid validation aren't just helpful, they're existential. For Volvo Cars, strategically navigating this AI frontier, this is where Dify delivers indispensable value."

Interpretation: the value lies in rapid validation capability, rather than deep customization.

Case 2: RICOH (General Manager of a Business Division)

"What makes Dify stand out is its ability to democratize AI agent development. By combining powerful AI/ML capabilities on a no-code platform, its rapid deployment and intuitive interface make it highly accessible even for beginners, significantly accelerating citizen development."

Interpretation: the core value is citizen development — enabling non-engineering roles to participate in building agents.

Case 3: ETS, an assessment products company

"Dify's intuitive interface allows our teams to rapidly design and deploy complex natural language processing pipelines, enabling us to improve the quality of our assessment products while reducing cost and time to market."

Interpretation: points to reduced time to market and cost, but no specific percentages are given.

Other customer logos listed on the official site: Maersk, Adobe, Google, Panasonic, PayPal, Lilly, Thermo Fisher Scientific, MITRE, Mercedes-Benz, Novartis, Deloitte, McDonald's, AIA, Volkswagen, CTC, KPMG, Volvo, RICOH, ETS, Anker, and more. Note: displaying a brand logo is not an endorsement of effectiveness, nor does it represent deployment scale.

Vendor-reported scale figures (not independently verified): 280+ enterprise customers; Dify runs on over 1.4 million machines in more than 175 countries and regions; over 5 million downloads; more than 1 million apps running on the platform. All of the above are vendor-published figures; this document records and labels them as-is, and they are not to be used as the basis for conclusions.

Part for which no public quantitative data was found: as of the search date 2026-09-12, no cost-reduction percentages, task success rates, labor substitution rates, or ROI data were found for any Dify customer case; this is stated truthfully here and not filled in.

Risk note (to be included in selection evaluation): in 2026, Dify's R&D team shrank from 40 to 15 people. This is a clear signal of organizational change; when evaluating long-term dependence, its impact on release cadence and enterprise support capability must be considered. This figure is verified project hard data; its specific background and scope of impact were not further researched.

7. Summary

7.1. Strengths

1. Strongest RAG: the most fully productized retrieval pipeline in this group, with parent-child chunking + reranking + Knowledge Pipeline.
2. Lowest low-code barrier: non-engineering roles can participate in building, with strong citizen development capability.
3. Deliverable capability across all six layers: the only platform in this group covering L1~L6 (L6 requires Enterprise).
4. Bidirectional MCP: can call the external ecosystem and publish its own capabilities as MCP tools.
5. Human Input node: design-time-fixed pause points, suited to regulated industries.
6. Self-hosting with no per-message billing: data self-control + controllable cost.
7. Large ecosystem: 155K+ GitHub stars, running on 1.4 million machines (vendor figures).

7.2. Weaknesses

1. License not OSI: prohibits multi-tenant commercial hosting and logo removal; building SaaS on top of it requires a commercial license.
2. Weak L4: no run-level checkpoints or artifact versioning; long tasks cannot resume after a crash.
3. Incomplete L5: no native Golden Dataset / regression set; evaluation relies on annotation and external tools.
4. Limited canvas expressiveness: complex logic degenerates into spaghetti wiring, reducing maintainability.
5. Pronounced L6 tiering: SSO / RBAC / audit / compliance all live in Enterprise; the Community edition has essentially none.
6. Dify Agent risk: the Beta version's official warning limits it to trusted users only.
7. Organizational change risk: R&D team reduced from 40 to 15 people (2026).

7.3. Applicable Boundaries

7.4. Selection Recommendations

• The trade-off with LangGraph is two ends of the same axis: Dify trades depth for speed of onboarding; LangGraph trades onboarding cost for control. For teams without strong platform engineering capability, Dify offers clearly better value for money.
• Like Coze (see 06-coze.md), it belongs to the Agent Platform category; the difference is: Dify is open source and can be privatized with a larger global ecosystem, whereas Coze relies on ByteDance's ecosystem and domestic channel distribution.
• The trade-off with n8n (see 07-n8n.md): choose Dify when it's 80% AI, and n8n when it's 80% pipeline integration.
• Three things must be confirmed before adoption: whether the license terms permit your business model; whether Enterprise SSO / audit fits your budget; whether your critical dependencies have a fallback after the R&D team contraction.

Information Gap Statement

1. Company headquarters: third-party sources respectively label San Francisco and New York, unresolved; marked [To be verified].
2. Latest version and release cadence: v1.16.1 (2026-07-28) comes from third-party verification, not compared item-by-item with the official release feed; marked [To be verified].
3. GitHub star count: the official site's 155K+ and third-party 138K / 153K+ differ by timing and methodology; the statistical date must be noted when referencing.
4. Enterprise pricing: third parties mention a $150,000 / 12 months figure on AWS Marketplace, not officially confirmed; marked [To be verified].
5. Quantified customer-effect data: none of the official-site customer statements include quantifiable metrics; not filled in.
6. Layoff background: the R&D team's drop from 40 to 15 people (2026) is verified project hard data, but its specific causes, scope, and impact on the product roadmap have no official explanation; marked [To be verified].
7. ARR data: third parties estimate about $3M ARR, with no official disclosure; marked [To be verified].
8. Knowledge Pipeline data-source list: compiled from third-party sources, not checked item-by-item against official documentation.

8. References

1. Dify official website — LangGenius, Inc. https://dify.ai/
2. Dify — AI Wiki (funding, versions, features, and license notes). https://aiwiki.ai/wiki/dify
3. Dify — Open-Source LLMOps Platform With Visual Workflow Builder — Hivebook (version verification and license terms excerpts). https://hivebook.wiki/wiki/dify-open-source-llmops-platform-with-visual-workflow-builder
4. Dify Review — Open-Source AI Workflow Platform With MCP, RAG, and Multi-Agent Orchestration — ChatForest (bidirectional MCP, Human Input node, 2026 updates). https://chatforest.com/reviews/dify-open-source-ai-workflow-agent-platform-review
5. Dify: Build LLM Apps Visually — Open-Source Guide (2026) — Diego Rodriguez (deployment commands, tech stack, license terms). https://www.diego-rodriguez.work/blog/dify-llm-app-builder-2026-guide
6. Dify in Practice: Full-Stack Low-Code Platform from RAG to Agent Workflows — AgentList (chunking strategies, deployment architecture, multi-tenancy). https://www.agentlist.top/en/articles/dify-ai-platform-practice
7. Dify — EveryDev.ai (capability matrix, deployment models, enterprise adoption signals). https://www.everydev.ai/tools/dify
8. LangGenius, Inc. company profile — ToolJunction (funding and valuation). https://www.tooljunction.io/ai-tools/dify-ai/company
9. Project parameter card v1.0 (Agent Platform definition and six-layer capability model) — this project's internal baseline document.
10. R01-overview research report (three-layer distinction of Framework / Harness / Platform) — this project's internal research report.