Overview

Relevant source files

• .github/workflows/tests.yml
• Dockerfile
• README.md
• README_ar.md
• README_fr.md
• README_id.md
• README_ja.md
• README_ko.md
• README_pt_br.md
• README_tr.md
• README_tzh.md
• README_zh.md
• api/apps/__init__.py
• api/apps/backward_compat.py
• api/settings.py
• api/utils/api_utils.py
• api/utils/validation_utils.py
• docker/.env
• docker/README.md
• docker/docker-compose-base.yml
• docker/infinity_conf.toml
• docs/guides/manage_files.md
• docs/quickstart.mdx
• docs/references/http_api_reference.md
• docs/references/python_api_reference.md
• docs/release_notes.md
• helm/values.yaml
• pyproject.toml
• sdk/python/pyproject.toml
• sdk/python/ragflow_sdk/modules/__init__.py
• sdk/python/ragflow_sdk/modules/agent.py
• sdk/python/ragflow_sdk/modules/base.py
• sdk/python/ragflow_sdk/modules/chat.py
• sdk/python/ragflow_sdk/modules/chunk.py
• sdk/python/ragflow_sdk/modules/dataset.py
• sdk/python/ragflow_sdk/modules/document.py
• sdk/python/ragflow_sdk/modules/session.py
• sdk/python/ragflow_sdk/ragflow.py
• sdk/python/uv.lock
• test/testcases/test_web_api/test_system_app/test_apps_init_unit.py
• uv.lock
• web/src/components/document-preview/md/index.tsx
• web/src/constants/markdown-remark-plugins.ts

Purpose and Scope

RAGFlow is an open-source Retrieval-Augmented Generation (RAG) engine that integrates deep document understanding with agentic workflow capabilities. It is engineered to transform complex, unstructured data into high-fidelity, production-ready AI systems by combining intelligent document parsing, embedding, and retrieval with configurable LLM-powered agents and workflows. README.md76-78 pyproject.toml4

Core Characteristics: RAGFlow specializes in parsing complex document formats such as PDF, DOCX, Excel, and PPT using state-of-the-art deep learning models for document layout and table extraction. It segments documents into meaningful chunks using multiple, template-driven chunking strategies, generates contextual embeddings for these chunks, and stores them in a hybrid vector and keyword index. It supports conversational AI applications backed by traceable, grounded citations. Additionally, RAGFlow includes a visual agent workflow system for creating multi-step AI agents with persistent memory and tool-calling capabilities. README.md112-122 pyproject.toml4 README.md99-100

Key Value Statement: "Quality in, quality out" — the system prioritizes deep document understanding to reliably find relevant information even from vast, unbounded token spaces. README.md114-118

Sources: README.md75-122 pyproject.toml1-8

System Architecture

RAGFlow is built upon a multi-tier microservices architecture, optimizing scalability and decoupling synchronous API operations from intensive background processing tasks. The architecture follows a producer-consumer pattern with Redis Streams managing task queues. docker/.env140-146 docker/.env20

High-Level Architecture Diagram

Sources: README.md141-145 docker/.env13-182 docker/docker-compose-base.yml1-250 Dockerfile1-160 docker/.env180-181

Three-Tier Architecture

RAGFlow is logically divided into three tiers:

1. Frontend/API Tier:

   • Exposes REST and SDK endpoints via a Quart server running in asynchronous mode. pyproject.toml145-147 pyproject.toml101-102
   • Supports a hybrid deployment mode (API_PROXY_SCHEME=hybrid) where a Go-based server handles high-performance native components (ingestion, sync, search) and the Python server handles logic. docker/.env183-184
   • Provides a Model Context Protocol (MCP) server for tool and context integration. docker/.env179
   • Integrates multiple chat channels (Feishu, Discord, Telegram, WhatsApp, WeCom, etc.) via a reconciliation loop. README.md91 docs/release_notes.md62-65

2. Asynchronous Task Tier:

   • Background workers implement document parsing, embedding, and indexing asynchronously.
   • Tasks are queued in Redis Streams and consumed by the TaskExecutor. docker/.env140-146
   • A dedicated sandbox executor manager provides isolated environments for Python/JS code execution using gVisor. docker/docker-compose-base.yml158-184 docs/quickstart.mdx35

3. Persistence Tier:

   • Relational Database: MySQL 8.0 (or PostgreSQL) for metadata, tenancy, and session management using the Peewee ORM. docker/docker-compose-base.yml186-212 pyproject.toml134
   • Document Store: Choice of Elasticsearch, Infinity, OpenSearch, OceanBase, or SeekDB for vector and keyword search. docker/.env13-20
   • Object Storage: MinIO for raw file persistence. docker/docker-compose-base.yml214-229
   • Cache/Queue: Redis (Valkey) for task orchestration and ephemeral state. docker/.env140-146 pyproject.toml122

Sources: pyproject.toml9-182 docker/.env13-184 docker/docker-compose-base.yml1-250

Core Components and Services

System Entity Mapping Diagram

This diagram associates high-level RAG concepts with their primary implementation entities in the codebase.

Sources: docker/docker-compose-base.yml158-161 docker/.env180-184 README.md116-120

Application Entry Points

Service/Subsystem	Main File(s)	Description
Python API Server	api/ragflow_server.py	Quart-based async server for REST API and SDK.
Go Server	cmd/ragflow_server.go	Native high-performance service layer (API/Admin/Ingestor/Syncer).
Task Executor Worker	rag/svr/task_executor.py	Background consumer for ingestion/chunking tasks (Python).
Sandbox Manager	sandbox-executor-manager (image)	GVisor-based isolated code execution service.
MCP Server	mcp/server/server.py	Model Context Protocol implementation.

Sources: docker/.env175-182 docker/docker-compose-base.yml158-184 docs/quickstart.mdx35

Key Features

• Deep Document Understanding: Extracts structured knowledge from complex formats (PDF, DOCX, Excel, PPT) using vision-based layout analysis and OCR, including SoMark and PP-OCRv6. README.md116-117 docs/release_notes.md43 docs/release_notes.md66
• Intelligent Chunking: Provides template-based chunking (Naive, Book, Law, Paper, Presentation, etc.) that is explainable and adjustable. README.md120-123 docs/references/python_api_reference.md165-180
• Hybrid Retrieval: Combines vector search with keyword-based BM25 and multi-stage reranking across multiple document stores. README.md138 docker/.env13-20
• Agentic Workflow: Visual "Canvas" for orchestrating complex AI tasks, supporting tool-calling, multi-turn reasoning, and persistent memory. README.md99-100 README.md93-94
• Data Source Connectors: Incremental syncing from Google BigQuery, Confluence, S3, Notion, Discord, and Google Drive. docs/release_notes.md40 README.md95-96
• Chat Channels: Out-of-the-box integration for Feishu, Discord, Telegram, Line, WhatsApp, DingTalk, and WeCom. README.md91 docs/release_notes.md62-65

Sources: README.md88-140 docs/release_notes.md1-100 docs/references/python_api_reference.md120-210

Technology Stack

Backend

• Languages: Python 3.13, Go, C++. pyproject.toml8 .github/workflows/tests.yml135
• Frameworks: Quart (Async Python), Peewee (ORM), Eino (Go Agent framework). pyproject.toml101-102 pyproject.toml134
• ML/AI: Support for OpenAI (GPT-5 series), DeepSeek (v4), Gemini (3 Pro), Anthropic, and local models via Ollama. pyproject.toml17-133 README.md91-99
• Infrastructure: Docker, Docker Compose, Kubernetes/Helm, and gVisor for sandboxing. docs/quickstart.mdx33-35 docker/docker-compose-base.yml1-10

Frontend

• Stack: React-based web interface with Ant Design and Radix UI. README.md3
• I18n: Support for 13+ languages including English, Chinese, Japanese, Korean, French, and Dutch. README.md8-18 docs/release_notes.md18

Sources: pyproject.toml1-182 README.md1-140 docs/quickstart.mdx28-36 docs/release_notes.md12-100