RAGFlow

The open source RAG engine based on deep document understanding helps enterprises build knowledge base question and answer systems.

knowledge management

https://ragflow.io/

Visit Website

In-depth Report

RAGFlow is an open source RAG (Retrieval Augmented Generation) engine built on deep document understanding, designed to provide streamlined RAG workflows for businesses and individuals of all sizes. The platform integrates cutting-edge RAG technology and Agent capabilities to help users build reliable knowledge base question and answer systems. RAGFlow supports technologies such as vector search, BM25 full-text retrieval, tensor retrieval, and advanced reordering, and provides data ingestion pipelines, visual workflows, and AI agent building functions. Currently, the number of GitHub stars exceeds 40,000, making it a popular project in the open source RAG field.
RAGFlow is developed and maintained by the Infraflow team, and the project’s open source address is https://github.com/infiniflow/ragflow. The core members of the team come from search engine backgrounds such as Sogou and Baidu, and have profound accumulation in the fields of natural language processing and retrieval. The core positioning of this product is to solve the pain points of enterprise knowledge management and AI application implementation. Traditional enterprises face challenges such as confusing data formats, poor retrieval results, and serious hallucination problems when building knowledge bases. RAGFlow uses deep document understanding technology to extract high-quality semantic information from complex unstructured documents, significantly improving the accuracy of question and answer.
RAGFlow provides complete data processing and question and answer system building capabilities: The data ingestion pipeline is one of the core highlights of the product. Users can configure data cleaning and processing processes, and support the import and preprocessing of multiple document formats. The system will automatically analyze the document structure, extract key information and convert it into vector representation, laying the foundation for subsequent retrieval. Multimodal retrieval technology is another core capability. RAGFlow also supports three search methods: vector search (Vector), full-text search (BM25), and tensor search (Tensor), and provides advanced re-ranking (Re-ranking) function to optimize search results. Users can flexibly choose or combine different retrieval strategies according to actual needs. This multi-channel recall design significantly improves the recall rate and accuracy of retrieval. AI agents and workflow capabilities allow users to build complex question-and-answer processes through a visual interface. The platform provides a variety of preset workflow templates such as financial analysis, legal case analysis, equipment maintenance guides, etc., which users can quickly reference and customize. RAGFlow also integrates MCP (Model Context Protocol) to support seamless connection with large language models. Two solutions, cloud and local deployment, meet the needs of different scenarios. The cloud version is suitable for quick verification and small and medium-scale deployment, while the local deployment version meets enterprise compliance requirements for data security and privacy. From the perspective of user experience, RAGFlow's advantages are open source, free, comprehensive functions, and detailed documentation. The challenge is that deployment requires certain technical capabilities (Docker and model environments need to be configured), the interface currently only supports English and Chinese, and some advanced functions require high-performance hardware support.
RAGFlow adopts an open source free + cloud paid business model: The open source version is completely free, and users can download the source code on GitHub and deploy it themselves. The open source version is fully functional and suitable for secondary development and customization by the technical team. The cloud version (RAGFlow Cloud) provides hosting services and is billed based on usage. The cloud version eliminates deployment and maintenance costs and is suitable for teams that want to quickly start projects. Specific pricing can be viewed at https://cloud.ragflow.io/. Compared with competing products such as Pinecone (pure vector database, higher fees), Weaviate (open source but relatively single function), Qdrant (open source vector engine), RAGFlow's advantage lies in its end-to-end RAG solution, and the open source version is extremely cost-effective.
Judging from feedback from GitHub and the technical community, RAGFlow has received more positive reviews. Developers appreciate its excellent document processing, high retrieval accuracy, and flexible workflow configuration. The open source community is active and updated frequently. Negative feedback mainly focuses on: the deployment documentation is not friendly to novices, local deployment has high hardware requirements, and performance needs to be optimized when dealing with large-scale data.
In the field of open source RAG, RAGFlow is currently one of the hottest projects. Unlike pure vector databases such as Pinecone, Weaviate, and Qdrant, RAGFlow provides a complete end-to-end RAG solution, from data ingestion to Q&A output. Competitors include: LangChain (application framework, RAG capabilities are relatively basic), LlamaIndex (data indexing tool, more developer tools), Milvus (enterprise-level vector database, positioned at a lower level).
Scenarios suitable for using RAGFlow: Enterprises build internal knowledge base question and answer systems, developers need to quickly build RAG applications, AI applications require reliable knowledge retrieval capabilities, and technical teams are capable of Docker deployment and operation and maintenance. Situations not recommended for use: users with no technical background at all, scenarios that only require simple keyword searches, and real-time systems with extremely high response delay requirements.
RAGFlow is a powerful open source RAG engine, especially suitable for technical teams to build enterprise-level knowledge bases and AI question and answer systems. It has outstanding advantages in document understanding and retrieval effects, and the free use of the open source version lowers the threshold for use. The main challenge is that deployment and operation require certain technical capabilities. If your team has technical prowess and a need for data privacy, RAGFlow is an option worth trying.

User Reviews

BitShark586

—

用 RAGFlow 搭了个合同审核助手，表格解析是真的稳，跨页表格居然能自动拼接，之前用别的方案拆得一塌糊涂。

Joyce_Russell_704

—

部署确实有点门槛，Docker Compose 拉起来倒是不难，但要配好外部 LLM 和 embedding 得折腾半天。第一次搞建议直接走 Cloud 版。

XS894PJ1Y

—

自托管太香了，一台 $40 的 Hetzner VPS 跑了 5 个客户的知识库，总共五万多份文档，每个月 API 费才几十块。跟托管服务比省太多了！

JTorres_77

—

说它低代码我觉得是个误解，你要是不知道 chunk overlap 是什么意思、embedding dimension 怎么调、retrieval top-K 怎么设，做出来的东西根本没法用。我前两个项目全栽在分块策略上了——第一次 chunk 太大丢失精度，第二次太小丢失上下文，前后折腾了两周才调好。

JackPerez

—

社区真的活跃，30K GitHub stars 不是白给的，提了个 Excel 解析的 bug，5 天就合并了修复。Discord 群里也经常有人分享部署经验。

侠客_5

—

可视化流水线比 LangChain 那些框架直观多了，给律所搭了 15000 个案卷的知识库，拖拖拽拽 4 小时就搞定了，包括调 chunk 参数和测试。换了以前用 LangChain 干这活至少 3 天起步，还得写一堆胶水代码。

Judy.CastilloQ

—

深度分块确实牛，不再是简单 500 token 一刀切，表头跟内容在一起、段落上下文不丢。实测检索准确率比 naive chunking 高了差不多 20%。

iساراعلیزاده_dev

—

100 页的 PDF 解析要 2-5 分钟，这个速度说实话有点慢。刚接入新客户的时候一万份文档灌进去跑了十几个小时，客户差点等不及。建议批量导入的时候安排好时间窗口，白天别搞大规模 ingestion。

BGonzales_2022

—

云端的定价有点离谱，$49 一个月才 100MB 存储空间，正经企业文档随便就超了。还是自托管划算。

KennethMurphy_q

—

ARM Mac 用户劝退，官方 Docker 镜像只有 x86 的，自己 build 踩了一堆坑。最后换了台 x86 服务器才算消停。

Keith.James_X

—

新版本改成 slim 镜像了，不内置 LLM 和 embedding，得自己配外部服务。好处是灵活了，坏处是部署步骤又多了两步。

Vincent_PhillipsJr00

—

四款开源知识库（FastGPT、WeKnora、Dify、RAGFlow）都测了一遍，RAGFlow 的文档解析是唯一一个能把复杂表格完整保留的，92% 的识别率，服气。其他三款在跨页表格和合并单元格上都不同程度地翻车了，对我这种做招标合同场景的来说没有别的选择。

SharonCarter_66

—

SSPL 协议要注意一下，如果你打算拿它做 SaaS 对外卖的话需要买商业授权。企业内部用完全没问题。

清风_26

—

我们是律所，主要用来检索历史判例和合同条款。答案带引用溯源这个功能太重要了，合伙人只认「哪个文件第几页说的」，RAGFlow 给得清清楚楚。之前试过 LangChain + Chroma 的方案，溯源功能得自己写，而且准确率远不如这个。

TuckerRoberts

—

跟 Dify 比的话，RAGFlow 在文档解析上强太多，但 Dify 的全栈应用编排能力更强。如果只做知识库选 RAGFlow，要做复杂 AI 应用就 Dify。

Brenda.PowellX

—

Agent 功能最近的更新挺惊喜的，能接 MCP、做可视化 workflow，把 RAG 和工具调用串起来。不过对纯做检索的人来说可能用不上。

angrymouse550

—

生产环境部署一定要先配好监控和备份，我有一次 Elasticsearch 挂了，整个知识库查不了，排查了半天才发现是 ES 配置的问题。

Gary_Hill

—

chunk 可视化这个功能太棒了，能直接看到文档是怎么被切的，不对的地方还能手动改。数据质量可控，安心很多。

MsMileKapetanović_2024

—

我们金融行业用着挺好，合规那边要求所有 AI 回答必须有据可查，RAGFlow 的 citation 功能完美满足。就是部署初期调参花了点时间。

LaurieCook

—

yyds，真正能处理扫描件里表格的 RAG 工具，没有之一。其他方案遇到扫描件的表格直接崩，它居然能准确还原行列结构。

Edward_Ross_7723

—

有个坑，vm.max_map_count 不改的话 Elasticsearch 根本起不来，我第一次部署卡在这步一个多小时。Linux 用户务必注意。

侠客_18

—

对比了一圈还是选 RAGFlow，FastGPT 表格解析太弱（合并单元格直接崩），Dify 没内置 OCR 得自己花钱买插件，WeKnora 不支持 Excel 和 PPT。只有 RAGFlow 全格式通吃，虽然不好装但值得，生产环境已经稳定跑了两个月了。

LKelly007948

—

说实话学习曲线是陡的，头两周基本在跟 chunk 策略和 embedding 模型较劲。但过了那个坎之后是真顺手，现在客户都很满意。

MichaelMitchell_99971

—

有个做医疗的朋友也用的 RAGFlow，他们要求数据绝对不能出内网，自托管完美满足合规。处理临床指南和文献的效果据说很好。

xcfj82

—

容量规划要做好，50GB 磁盘是最低要求，实际跑起来索引文件涨得很快。我一开始只分了 80GB，三个月就不够了。

DennisEdwards_202029

—

Text2SQL 功能挺惊喜的，可以直接用自然语言查数据库，虽然还不够完美但已经能解决不少日常查询需求了。

Brittany.Perry52083

—

DeepDoc 引擎确实是核心卖点，普通的 RAG 工具只能提取文字，它能理解文档的版面结构。我试过一份带三层嵌套表格的财报 PDF，完美解析，连单元格里的合并注释都识别出来了，换别的工具早成一团乱码了。

墨染445

—

开了重排序之后效果提升明显，BGE-Reranker 一上，top-5 结果的相关性高了一个档次。建议必开。

Catherine.White_88

—

提醒一下，GPT-4 做 LLM 的话 API 费用不低，我后来换了 DeepSeek V3 成本降了 90%，效果差不太多。

happywolf343

—

个人用的话建议先上 Cloud 免费版试试水，虽然免费版只给 0.1GB 存不了多少文档，但功能体验是一样的。觉得行再自己部署。

RAGFlow

In-depth Report

User Reviews

Similar Products