The open source RAG engine RAGFlow continues to evolve: it supports DeepSeek v4 and multi-platform access, and local knowledge base deployment enters the agent-based stage.

News Highlights

• RAGFlow is an Apache-2.0 open source retrieval-enhanced generation (RAG) engine. GitHub shows 88.3k stars, 10.4k forks, and 737 contributors, maintaining high activity.
• Intensive recent updates: Support DeepSeek v4 from 2026-04-24; support access to multiple chat channels such as Feishu, Discord, Telegram, and Line from 2026-06-15.
• The focus of capabilities shifts to Agentization: AI Agent Memory is introduced on 2025-12-26, and Agentic Workflow and MCP are supported on 2025-08-01.
• The local deployment threshold is clear: the official prerequisites are CPU ≥ 4 cores, memory ≥ 16GB, disk ≥ 50GB, and Docker ≥ 24.0.0.

Dismantling of key information

Project and model support

• Positioning: RAG workflow for enterprises of any size, focusing on in-depth document understanding of "Quality in, quality out".
• Wide model adaptability: Supports OpenAI GPT-5 series, DeepSeek v4, Gemini 3 Pro, etc., and can configure LLM and Embedding models locally or in the cloud as needed.

Core Competencies

• Template Chunking: Intelligent segmentation that is interpretable and capable of manual intervention.
• Evidenced Reference: Visualized segmentation results + traceable references to reduce illusions.
• Heterogeneous data sources: Support Word, PPT, Excel, TXT, scanned documents, web pages, etc.
• Multiple Recall + Fusion Rearrangement: Automated RAG arrangement, taking into account both individuals and large enterprises.

Data access and channels

• Support data synchronization from Confluence, S3, Notion, Discord, Google Drive, etc. (from 2025-11-12).
• Supports document parsing methods such as MinerU and Docling (from 2025-10-23).
• Multiple chat channels: Feishu, Discord, Telegram, Line (starting from 2026-06-15).

Local deployment

• The current version of the official Docker image is v0.26.4 (subject to the official Releases real-time page).
• Prerequisites: CPU ≥ 4 cores, memory ≥ 16GB, disk ≥ 50GB, Docker ≥ 24.0.0 and Docker Compose ≥ 2.26.1; Elasticsearch is used by default to store full text and vectors, and can be switched to Infinity.

Impact Analysis

• Enterprise knowledge base moves from "capable of answering questions" to "capable of doing work": Agent Memory + MCP makes the RAG knowledge base not only a question and answer database, but also capable of participating in complex tasks.
• Domestic model linkage: Support for DeepSeek v4 and other models gives Chinese companies more choices in cost and data compliance.
• Open source ecosystem accelerates involution: Open source solutions such as RAGFlow, Dify, and AnythingLLM compete with cloud solutions, and the threshold for localized deployment continues to decrease.

Practical Advice

• Enterprises with data compliance requirements give priority to evaluating local deployment, and Docker Compose can be launched for trial use with one click.
• If the team size is small, you can start with RAGFlow official Cloud or stand-alone Docker, and verify the results before moving to production.
• The production environment must plan the segmentation strategy, recall rearrangement and permissions to avoid "the library is difficult to use when it is large".

Reference sources

• RAGFlow GitHub: https://github.com/infiniflow/ragflow
• RAGFlow official website: https://ragflow.io
• RAGFlow documentation: https://ragflow.io/docs/dev/