RAGFlow

1. Project/Product Overview

2. What does it mostly do?

The core positioning of the RAGFlow is " RAG Agent integrated context engine " -- not only does document retrieval, but also provides a set of full-link capabilities from document understanding to intelligent Agent.

Core Architecture (Six Modules)

用户/API 层  →  Chat / Agent / 搜索界面 / REST API
     ↓
Agent 编排层 →  Workflow（人工编排） + Agentic Workflow（LLM 自主规划）
     ↓
检索层     →  多路召回（向量 + 关键词 + 知识图谱） + 融合重排序
     ↓
索引层     →  Elasticsearch / Infinity（全文 + 向量混合索引）
     ↓
知识提取层  →  DeepDoc 引擎 / MinerU / Docling / OpenDataLoader
     ↓
数据源层    →  文件上传 / 数据源连接器（Confluence、S3、SharePoint、飞书等）

Main function module

3. Applicable Scenario

4. Not quite the scene

5. Core Competence List

5.1 Document parsing capability (DeepDoc engine)

5.2 block strategy

5.3 Agent / Workflow Capability

5.4 Models and Infrastructure

5.5 enterprise-class features

6. Architecture/deployment/integration approach

Deployment Mode

Deployment steps (Docker)

# 1. 确保 vm.max_map_count >= 262144
sudo sysctl -w vm.max_map_count=262144

# 2. 克隆仓库
git clone https://github.com/infiniflow/ragflow.git
cd ragflow/docker

# 3. CPU 模式启动
docker compose -f docker-compose.yml up -d

# 4. GPU 模式启动（需 NVIDIA GPU）
# sed -i '1i DEVICE=gpu' .env
# docker compose -f docker-compose.yml up -d

# 5. 检查状态
docker logs -f docker-ragflow-cpu-1

# 6. 浏览器访问 http://<服务器IP>

LLM Integration Example (configured in WebUI)

Supports access to any model through OpenAI-compatible APIs:

-Ollama:'http://host.docker.internal:11434/v1' (local model)

-vLLM:'http://vllm-server:8000/v1'

-Alibaba Cloud Tongyi Thousand Questions: DashScope API Key

-DeepSeek: by DeepSeek the API Key

-Baichuan/ZhipuAI/MiniMax / Moonshot: respective API Key access

How to use #7.

Docker deployment WebUI usage (recommended way to get started)

# 1. 部署（见上节）
git clone https://github.com/infiniflow/ragflow.git && cd ragflow/docker
docker compose -f docker-compose.yml up -d

# 2. 浏览器登录 http://localhost 或 http://<服务器IP>
#    首次登录需注册管理员账号

WebUI Usage Process:

1. Configure Model Provider: Go to the "Model Provider" page and add LLM (e. g., the common qwen-plus), embedded model (e. g., BGE), and reorder model.
2. Create Knowledge Base: Click "Knowledge Base" → "New", and configure the chunking strategy (such as General) and PDF parser (DeepDoc) after naming it.
3. Upload Document : Drag or select PDF/Word/PPT/Excel/image file, RAGFlow automatically parse and index
4. Configure Search: Select the mixed search mode (vector keyword) and set the Top-K and similarity thresholds.
5. Start Dialogue : Switch to the "Dialogue" tab and enter questions to get answers with traceability.
6. Build Agent: Enter the "Agent" page, use the preset template or build from the blank canvas, and drag and drop the component to arrange the process.
7. Release Channel : Release Agent as Bot for channels such as flying book/Discord/Slack, or embed its own system through API

Python SDK Usage

from ragflow_sdk import RAGFlow

# 连接 RAGFlow
rag = RAGFlow(
    api_key="your_api_key",
    base_url="http://localhost/v1"
)

# 创建/获取知识库
dataset = rag.create_dataset(name="企业规章制度")

# 上传文档
dataset.upload_documents(["员工手册.pdf", "薪酬制度.docx"])

# 检索
chunks = dataset.retrieve(
    question="年假怎么请？",
    top_k=5
)

# chat
for ans in dataset.chat_stream("年假怎么请？"):
    print(ans, end="")

REST API calls

curl -X POST http://localhost/v1/api/chats \
  -H "Authorization: Bearer " \
  -H "Content-Type: application/json" \
  -d '{
    "dataset_ids": [""],
    "question": "请说明年假申请流程",
    "stream": true
  }'

8. What can I say before sales

8.1 a sentence positioning

• * "RAGFlow is the most popular RAG open source engine on GitHub in the world -- 84,000 Stars,DeepDoc the strongest in-depth document understanding capability in the industry, allowing AI to real' read' your PDF, contracts and reports. "**

8.2 customer pain points → solutions

8.3 Differentiated Selling Points

vs MaxKB(1Panel Product):

vs Haystack(deepset products):

vs LlamaIndex:

vs Dify:

Core differences in a word:

- You need to build a document knowledge base, process PDF/contract/scanned documents → RAGFlow (document understanding is invincible)

- You want to build complex AI applications, Workflow → Dify (more mature choreography)

-You want code level fully controllable → Haystack / LlamaIndex

-You want simple FAQ lightweight → MaxKB

8.4 Customer Value Story Line

1. Cut in (cause pain) :"Have you tried to use RAG system to process contract/research report, and AI of the results mixed up all the form data?"
2. Dismantling (Root Cause of Location) :"The root cause is not that LLM can't do it, but that document parsing can't do it-traditional RAG cuts PDF as plain text, tables are cut into pieces, and naturally cannot be retrieved."
3. Demonstration (muscle show) : Upload a complex PDF (including scanned pages of tables and charts) on the spot → DeepDoc analysis → accurate question and answer → trace the source to the original location. This is the most powerful demo.
4. Extension (from RAG to Agent) :"Not only Q & A. You can drag and drop components to build Agent-automatically read contracts → extract key terms → compare historical cases → generate analysis reports."
5. Landing (dispelling worries) :"Docker is deployed to your intranet with one click, and the data does not go out. API to interface with existing systems. Flying book/DingTalk robot, direct use by employees."
6. End (Trust evidence) :"84,000 developers worldwide Star RAGFlow. Alibaba Cloud official recommendation. Open source free, Apache-2.0 protocol, no risk."

9. Frequently Asked Customer Questions

10. PoC Recommendations

Recommended PoC Direction: RAG System for Complex Document Knowledge Base

Total: about 4 working days

Recommended validation metrics:

-Retrieval recall rate (including table/image content)> 85%

-complex table data question and answer accuracy> 90%

-with traceability reference answer ratio> 95%

-End-to-end average response time <5 seconds

-Scan OCR accuracy> 95%

Key to the success of PoC:

-Must use customer real complex documents-DeepDoc ability to process these documents is the core selling point

-Focus on demonstration form question and answer and scan analysis, which is the place where the gap can be widened best.

-The traceability function must be displayed-click on the answer to jump to the original document position.

-GPU acceleration should be matched well, otherwise slow DeepDoc parsing speed will affect the demonstration experience.

11. Risks and Considerations

12. My Pre-Sales Judgment

Recommended: Most strongly recommended (RAGFlow is preferred for customers who need a document knowledge base/complex document RAG)

Reason:

1. Document understanding barrier is the highest : OCR + TSR + DLR three-in-one vision model of DeepDoc engine, complex form F1 > 95%, which is the RAGFlow's hardest moat. Other frameworks (Dify, MaxKB, Haystack) have significant gaps in this area and are difficult to catch up in the short term.
2. Global Community First :84,059 Stars not only represents recognition, but also means rich community resources-rapid problem solving, best practice precipitation and rich third-party integration.
3. Platform Complete : From Document Parsing → Block → Retrieval → Reordering → Generation→Agent→Channel Publishing, Full Link Coverage and Can be Operated through WebUI. Not the framework, the product.
4. Enterprise Friendly :Apache-2.0 is completely free for commercial use + Docker one-click deployment + data not going out + multi-channel access + Admin management plane. It is very friendly to Chinese enterprise customers (especially the needs of Xinchuang/localization).
5. Fast iteration speed :7,000 + Commits, a major version every 1-2 months. MCP, GraphRAG, Agent, Memory and other cutting-edge capabilities are quickly followed up.
6. Alibaba Cloud Endorsement: Recommended by Alibaba Cloud SAE official integration, highly available instances can be deployed with one click of SAE.

Recommended Customer Persona:

-There are a large number of complex documents that need RAG processing (PDF contracts, scans, reports with tables/charts)

-Need to secure data through private deployment

-Chinese scene-based

-Requires low code/visual operations (not pure code framework)

-Limited budget but pursue RAG effect (open source free)

-There are compliance requirements for Xinchuang/localization

-Requires multi-agent collaboration MCP tool integration

Not recommended situations:

-Requires strong Workflow orchestration capability (branch/loop/variable management) → Dify

-Pure code framework, Python developers who need extreme flexibility → Haystack / LlamaIndex

-Sufficient budget and need international large factory enterprise version support → Haystack Enterprise

-Simple FAQ, lightweight requirements, low hardware budget → MaxKB / FastGPT

-ARM64 platform and unable to build images on its own

-Overseas model ecology is the core demand (Dify's 56 providers are more abundant)

13. REFERENCE

-GitHub repository: https://github.com/infiniflow/ragflow

-Official Document: https://ragflow.io/docs/dev/

-Official website: https://ragflow.io

-SaaS Cloud Edition: https://cloud.ragflow.io

-Release Notes (Release Notes):https://ragflow.io/docs/dev/release_notes

-DeepDoc technical details: https://github.com/infiniflow/ragflow/tree/main/deepdoc

-Roadmap:https://github.com/infiniflow/ragflow/issues/12241

-Discord Community: https://discord.gg/NjYzJD3GM3

-Python SDK:https://pypi.org/project/ragflow-sdk/

-Helm Chart:https://github.com/infiniflow/ragflow/tree/main/helm

-Agentic Workflow details: https://www.ragflow.io/blog/agentic-workflow-whats-inside-ragflow-v0.20.0

-Alibaba Cloud SAE deployment: Alibaba Cloud SAE supports one-click deployment of RAGFlow highly available instances

-CSDN DeepDoc Technical Analysis: https://blog.csdn.net/wayle123/article/details/159760654

• Analysis Date: 2026-07-02 | Data Aging: GitHub Metadata Pull in Real Time, Product Functions Based on v0.26.2 Official Documents and Release Notes *