RAGFlow is an open-source Retrieval-Augmented Generation engine that fuses RAG with agent capabilities into a context layer for large language models, built for developers and teams that need to turn complex, unstructured documents into production AI systems.

What it is

RAGFlow is an Apache-2.0 licensed project published under the AI & Machine Learning / AI Development Platforms category in this registry, listed here with 90,881 stars and 10,757 forks, and described as a Go project in the registry metadata while its self-hosting path also requires Python. At its centre sits a converged context engine paired with pre-built agent templates, so developers can move from raw documents to a working AI system without assembling the pipeline themselves. The engine is built around deep document understanding, template-based chunking, and grounded citations, and it exposes an API surface for integration into business applications.

The concrete problem it solves is the hand-wired retrieval stack. Teams that build RAG themselves must otherwise stitch together document parsers, chunkers, embedding models, vector recall, re-ranking, and prompt assembly, then keep that glue working as models change. RAGFlow replaces that bespoke orchestration with a single workflow: configurable LLMs and embedding models, multiple recall paired with fused re-ranking, and chunking that can be inspected and corrected by a human before answers are generated. It lives in the same ecosystem as the LLM providers it consumes, currently including DeepSeek v4, Gemini 3 Pro, and the GPT-5 series.

Key capabilities

• Deep document understanding via the deepdoc module, extracting knowledge from unstructured data in complicated formats.
• Template-based chunking with visualization of the resulting chunks, so a person can intervene before ingestion is finalised.
• Grounded citations with traceable references, aimed at reducing hallucination in generated answers.
• Compatibility with heterogeneous sources: Word, Slides, Excel, TXT, images, scanned copies, structured data, and web pages.
• Configurable LLMs and embedding models, multiple recall, and fused re-ranking in one orchestration path.
• Agentic workflow with MCP support, plus a Python/JavaScript code executor component that requires gVisor.
• Data synchronisation connectors for Confluence, S3, Notion, Discord, and Google Drive, and multiple chat channels including Feishu, Discord, Telegram, and Line.

Who uses it and how

• Enterprises of any scale running internal knowledge search where answers must carry traceable citations back to source documents.
• Application developers integrating retrieval into business software through the project's APIs rather than a separate search service.
• Teams building agent workflows that need agent Memory, MCP tool calls, and sandboxed code execution.
• Organisations with large existing document estates that want Connector-based sync from Confluence, Notion, Google Drive, or S3 instead of manual uploads.
• Self-hosting operators running the Docker and Docker Compose deployment on machines with at least 4 CPU cores, 16 GB RAM, and 50 GB disk.

Getting started

A hosted cloud option is available at https://cloud.ragflow.io. Self-hosting follows the repository's Docker and Docker Compose instructions after raising vm.max_map_count to at least 262144 and installing Python 3.13 or newer.

How it compares

The facts provided name no paid products that RAGFlow replaces, and they name no comparable open-source retrieval engine either. On the record available here, it stands alone in this registry.

When to use it — and when not to

A self-hoster must operate the Docker and Docker Compose stack, tune vm.max_map_count, meet the 4-core, 16 GB RAM and 50 GB disk floor, and install gVisor separately if the code executor sandbox is wanted — although the README excerpt does not spell out the backing database, object storage, or SMTP settings, so those must be confirmed in the full documentation before committing. Anyone unwilling to run that infrastructure should take the hosted cloud service instead. The listing also carries 1,525 open issues, which is a large backlog to weigh against the project's recent activity.