RAGFlow
RAGFlow (Retrieval-Augmented Generation Flow) is an open-source RAG engine developed by InfiniFlow that provides a complete pipeline for document ingestion, semantic retrieval, and answer generation. Launched in April 2024, it has grown to approximately 82,000 GitHub stars with over 6,000 commits and more than 590 contributors (462 named and 128 anonymous, as of June 2026).[^c9] The project is licensed under Apache 2.0 and has received over 1,200 code contributions from the community.[^c3] The latest stable release is v0.26.2, published June 26, 2026, adding Chat Channels for WhatsApp, DingTalk, and WeCom alongside PP-OCRv6 fallback support and expanded internationalization.[^c23]
RAGFlow is evolving from a human-facing RAG engine into a data foundation layer for AI agent harnesses, operating at the persistent state layer while agent brains remain stateless.[^c20] This strategic direction positions RAGFlow as a context engine for agents, providing high-quality data across heterogenous sources through a unified retrieval interface. The company has targeted mid-2026 for the v1.0 release, featuring a Go-language rewrite of the online retrieval layer to handle the projected orders-of-magnitude increase in agent search requests.[^c22] A key insight driving this evolution is that RAG's primary users are shifting from humans to AI agents, fundamentally changing retrieval behavior and infrastructure requirements.[^c21]
The system addresses two core enterprise challenges: extracting usable information from complex unstructured documents and improving the reliability of large language model responses. Its architecture combines a vision-based document parsing module (DeepDoc), a hybrid retrieval engine, and a visual agent workflow orchestrator. The system operates through a three-tier microservices architecture: a Frontend and API tier exposing REST and SDK endpoints via a Quart async server and a Go server for native components, an Asynchronous Task tier using Redis Streams and NATS for task queuing with Python TaskExecutor and Go Ingestor workers, and a Persistence tier using a polyglot approach with MySQL or PostgreSQL for metadata, pluggable document engines including Elasticsearch, Infinity, OpenSearch, and OceanBase for vector and keyword indexing, MinIO for object storage, and Redis for task orchestration and caching.[^c17][^c18] The Go workflow framework Eino powers the agent canvas engine, and ONNX Runtime provides local deep learning inference for document layout and OCR.[^c19] DeepDoc performs layout recognition across ten component types including text, titles, figures, tables, headers, footers, references, and equations, along with OCR supporting 15 or more languages and table structure recognition for complex layouts.[^c5] Document chunking uses templates tailored to document types such as legal, research, and resumes, with visual inspection allowing human correction.[^c7] The retrieval pipeline incorporates an iterative refinement loop that automatically adjusts queries when initial context is insufficient, reducing hallucinated responses.[^c6] The hybrid retrieval engine fuses vector search with full-text BM25 search, achieving 95% recall rates with P99 latency under 800 milliseconds on one-million-document datasets. The technology stack uses a Python 3.13 backend with a React and TypeScript frontend, Elasticsearch or Infinity for vector storage, MinIO for object storage, Redis for caching, MySQL for metadata, and LiteLLM for integration with over 100 LLM providers.[^c8] Asynchronous task processing is handled through a custom Redis Streams-based task executor, and the InfiniFlow team of approximately 10 to 15 developers simultaneously develops both RAGFlow and the Infinity database engine.[^c10]
RAGFlow was named among GitHub's fastest-growing global open-source projects in 2025. It has been adopted by thousands of enterprises across finance, manufacturing, healthcare, and education, with documented reductions in manual compliance review workloads of up to 70% and equipment diagnosis times dropping from 45 minutes to 8 minutes in manufacturing contexts.[^c2] The project maintains a biweekly release cadence with the v0.26.x series adding Chat Channels for WhatsApp, DingTalk, WeCom, Discord, Feishu, Telegram, and QQ bot integration, auto-populated model lists, multiple API keys per provider, seven enterprise data source connectors (Outlook, OneDrive, Microsoft Teams, Slack, SharePoint, Salesforce, and Azure Blob Storage), GraphRAG indexing checkpoint and resume, and ongoing Python-to-Go migration of the agent canvas engine with 22 components.[^c16][^c14][^c23]