Newsletter, guides

Dify vs RAGFlow vs Onyx: Batteries-included RAG platforms

Dify vs RAGFlow vs Onyx compared for 2026. What each open-source RAG platform does, its license, its hardware needs, and which one fits your team.

, 9 min read, updated

You spent a week on Python scripts, Dockerfiles and a React chat window. The result is a RAG demo that answers questions over a folder of PDFs. It is nowhere near a product: no login, no document sync, no logs, no admin screen. With a code framework like LangChain or LlamaIndex, the retrieval logic is the small part. Everything around it is the week.

A batteries-included RAG platform is a self-hosted application that ships a full RAG product by bundling ingestion, retrieval, chat UI, API and admin tools in one deploy. Dify, RAGFlow and Onyx are three open-source options teams often compare. This guide answers the Dify vs RAGFlow question first, then places Onyx, and shows when you should still write code.

Key takeaways

  • Dify is a visual LLM app builder with workflows, agents, a knowledge pipeline and a published API, under a modified Apache 2.0 license.
  • RAGFlow is a retrieval engine built around deep document parsing (layout analysis, OCR and table recognition) with chunk-level citations, under Apache 2.0.
  • Onyx is a workplace search and chat platform that indexes 50+ company apps, with an MIT-licensed Community Edition and a paid Enterprise Edition.
  • Choose Dify to build many AI apps fast, RAGFlow when complex PDFs break your retrieval, and Onyx for chat over your company’s own tools.
  • Write code with LangChain, LlamaIndex or Haystack when your logic does not fit a platform’s nodes, or when RAG is one feature inside a larger product.

What is the problem with a week-long RAG demo?

The retrieval chain is a small part of a production RAG app. The rest is plumbing every team rebuilds: parsers for each file type, sync jobs, a vector index, a chat UI, an API, auth, logs and an admin view.

Flow from source documents through parsing, chunking, indexing and retrieval to the LLM, with the API, chat UI, auth and logs that a batteries-included platform ships around that core

Batteries-included platforms ship most of that plumbing in one deploy. You trade some control for weeks of saved work. The three platforms below make that trade in different places.

What is Dify?

Dify is an open-source platform for building LLM apps on a visual canvas, with RAG as one of its building blocks. You connect nodes for models, knowledge retrieval, tools and logic, and every app you publish gets a web UI and an API.

  • What it ships: a workflow canvas, a prompt IDE, and agents that call Marketplace tools, MCP servers or your own APIs.
  • Models: many hosted providers, plus self-hosted models behind an OpenAI-compatible server.
  • RAG features: a knowledge pipeline with general, parent-child and Q&A chunk structures, and keyword, vector or hybrid retrieval.
  • Operations: logs and monitoring in the app, with integrations for Langfuse, Opik and Arize Phoenix.
  • Hosting: Dify Cloud, or self-host with Docker Compose on at least 2 CPU cores and 4 GiB of RAM.
  • License: the Dify Open Source License, which is Apache 2.0 plus two conditions. You may not run it as a multi-tenant service without written permission. If you use its frontend, you must keep Dify’s logo and copyright notices.

The limit is the node catalog. When your logic needs something no node or plugin provides, you write a code node, call your own API, or move to a code framework.

What is RAGFlow?

RAGFlow is an open-source RAG engine that focuses on getting the right content out of messy documents. It parses layout, tables and scanned pages before chunking, so retrieval works on structured content instead of raw text runs.

  • What it ships: DeepDoc parsing (layout analysis, OCR and table recognition), template-based chunking you can inspect, and answers with traceable citations to the source chunks.
  • Beyond retrieval: an agent builder with agentic retrieval modes, and an MCP endpoint so external agents can query your datasets.
  • Knowledge compilation: an optional step in the 1.0 ingestion pipeline. An LLM builds artifacts such as a knowledge graph, a page index or a timeline.
  • Hosting: Docker Compose, with a recommended start of 4 CPU cores, 16 GB of RAM and 50 GB of disk.
  • License: Apache 2.0. As of October 2026, v0.27 is the stable line and v1.0 is in release candidate.

A knowledge graph is not built on every upload. You add it to a pipeline on purpose, and it costs LLM calls at ingest time. The default gain comes from parsing: a table stays a table, so a question about a quarterly figure can reach the row that holds it.

RAGFlow ingestion pipeline from document upload through the parser, the chunker, an optional compiler that builds a knowledge graph or page index, and the indexer, then retrieval with citations

What is Onyx?

Onyx is an open-source workplace search and AI chat platform that connects to your company’s apps and keeps their content indexed. It calls itself the context layer for your team and its AI agents.

  • What it ships: chat with major LLM providers or self-hosted models, and agentic RAG over a hybrid keyword and vector index.
  • Agents: custom agents with their own knowledge scope, deep research, web search and actions through MCP.
  • Connectors: sync from 50+ applications, including Slack, Confluence, Google Drive, Jira and GitHub, so the index stays current without upload scripts.
  • Permissions: connectors can be public or private. Syncing each source’s own access rules is an Enterprise Edition feature for a listed set of connectors.
  • Hosting: Docker, Kubernetes or Helm, or Onyx Cloud. A Lite mode runs the chat UI in under 1 GB of memory but cannot index documents.
  • License: the Community Edition is MIT. The Enterprise Edition adds SSO, RBAC, analytics and white-labeling.

Onyx is the product your employees use, not a backend you build on. If the goal is “ask questions across our wiki, tickets and Slack”, it is the shortest path.

Dify vs RAGFlow: What is the difference?

Dify is an app builder that includes RAG. RAGFlow is a RAG engine that includes some app building. Pick Dify when you need many apps, workflows and agents fast. Pick RAGFlow when answers fail because your documents are complex.

Platform Built for Strongest at License
Dify Building LLM apps and workflows Visual canvas, agents, published API Apache 2.0 plus conditions
RAGFlow Retrieval over complex documents Layout and table parsing, citations Apache 2.0
Onyx Search and chat over company apps 50+ connectors and sync MIT (CE), commercial (EE)
LangChain, LlamaIndex Custom code Control of every step MIT

You can also run both. RAGFlow does parsing and retrieval, and Dify or your own service calls its API for the app layer. For the code-first side of this choice, see LangChain vs LlamaIndex vs Haystack.

What should you compare when evaluating open-source RAG platforms?

Compare them on your own documents and questions, not on feature lists. Five checks settle most evaluations:

  1. Retrieval quality on your files. Load 20 real documents and 50 real questions. Measure whether the right chunk comes back, as in tuning retriever top-k on your own data.
  2. Parsing of your worst file. Upload the scanned PDF or table-heavy report that breaks your current setup, and read the chunks.
  3. License fit. Dify’s multi-tenant condition matters if you plan to offer it as a hosted service. MIT and Apache 2.0 do not carry that limit.
  4. Data sources. If content lives in SaaS tools, Onyx’s connectors save the most work. If it lives in files, all three accept uploads.
  5. Operations. Check hardware needs, upgrade paths, tracing and how each one handles permissions.
Decision flow that routes company app search to Onyx, complex PDFs and tables to RAGFlow, fast app building to Dify, and custom agent logic to a code framework

How do you choose for your use case?

Match the platform to the constraint that hurts most today. Four common cases each have a clear first pick.

A chatbot over product PDFs for the sales team, live tomorrow. Use Dify. Create a knowledge base, attach it to a chat app and share the web link. You get a UI, an API and logs without writing a backend.

A LangChain bot that fails on long financial reports. Use RAGFlow. Your problem is retrieval quality. Layout-aware parsing keeps tables and sections intact, and citations show which chunk each answer came from. For questions that chain several facts, add multi-hop query decomposition or try the knowledge graph artifact.

Company-wide chat over Slack, Confluence and Drive, on your own servers. Use Onyx. Its connectors keep the index in sync, and self-hosting keeps the data inside your network.

An agent that queries SQL, calls internal APIs and posts to Slack. Write code with LangGraph or another agent framework. Your logic is the product, so you need full control. The build a team of AI agents lab walks through that design.

What to do next

  1. Write down the one constraint that hurts most: time to launch, retrieval quality, or data spread across SaaS tools.
  2. Collect 20 representative documents and 50 real questions with known answers.
  3. Self-host the matching platform with Docker Compose and load the same files.
  4. Score retrieval hits and answer quality on your questions, and read the chunks of your hardest file.
  5. Check the license against your deployment plan before anyone builds on it.
  6. Wire tracing into the winner so you can see slow or wrong answers, as in Langfuse and Grafana monitoring.

Frequently asked questions

Is Dify or RAGFlow better for RAG?

RAGFlow is the better fit when your documents are complex, and Dify is the better fit when you need to build apps fast. RAGFlow parses layout, tables and scans before chunking and cites the source chunk for each answer. Dify gives you a visual canvas, agents and a published API for each app. You can also use RAGFlow for retrieval and call it from Dify.

Can Dify and RAGFlow be used together?

Yes. RAGFlow handles parsing, chunking and retrieval and exposes datasets through its HTTP API and an MCP endpoint. Dify handles the app layer: workflows, agents, the chat UI and the public API. Your Dify app calls RAGFlow for context instead of using its own knowledge base. Measure the latency of that extra network hop on your own traffic before you commit.

Are Dify, RAGFlow and Onyx free to self-host?

All three have free self-hosted editions, with different license terms. RAGFlow uses Apache 2.0. Onyx Community Edition uses MIT, and SSO, RBAC and permission sync need the paid Enterprise Edition. Dify uses Apache 2.0 with two conditions: no multi-tenant service without written permission, and keep its logo when you use its frontend. Your running costs are servers and model API calls.

Should I use Dify or build RAG with LangChain?

Use Dify when shipping a complete app quickly matters more than custom logic. Use LangChain when RAG is one feature inside a larger product, or when your retrieval and reasoning steps do not fit Dify’s nodes. Dify gives you the UI, API and logs on day one. LangChain gives you control of every step, and you build the rest yourself.

Which open-source platform gives a full self-hosted AI stack?

None of the three covers model serving, RAG, apps and model tracking in one box. Dify and Onyx both cover RAG, chat, agents and an API, and both can call self-hosted models. You still run model inference separately, for example behind an OpenAI-compatible server. Dify adds observability through integrations such as Langfuse, while Onyx focuses on search across company apps.

Sources

Build what you just read, with Param, live

  1. You read the guide

  2. Free Live LabBuild a Second Brain with Claude CodeA second brain over your notes, docs and Linear that answers with sources and serves the same memory to any agent over MCP.Free seat
  3. ProWatch the video coursesDesign a Software Factory for Your Coding Agents, Build Your First Software Factory Execution Harness and each new course as the sessions are cut. The first lesson of each is free after you sign in.$149 a year
  4. WorkshopAI Agents in Production for EngineersA multi-agent system deployed with an identity per agent, permissions, guardrails, an LLM gateway, traces and evals, plus a checklist for your own agents.$497
  5. BuildcampAgentic AI Buildcamp for Engineers & PMsDesign, build and deploy a production agentic AI system in four weeks$997

Subscribe to Top 1% Engineers

One issue at a time on building AI systems that hold up in production: architecture, the open source that matters, and what to do on Monday. Every number cited.

We email you each new issue. You can leave at any time. Privacy

More guides

See the newsletter and every guide