infiniflow/ragflow

An open-source Retrieval-Augmented Generation engine that combines deep document understanding with agent capabilities to build a context layer for LLMs.

What it is

RAGFlow is a self-hostable RAG engine aimed at developers and enterprises that need question-answering over their own documents. It extracts knowledge from unstructured data in complicated formats, chunks it using explainable templates, and pairs retrieval with re-ranking so that answers come back with traceable citations. Beyond plain RAG, it ships pre-built agent templates and an orchestration layer, positioning itself as a "context engine" that turns complex data into production AI systems.

Key features

Deep document understanding for knowledge extraction from unstructured data with complicated formats, including Word, slides, Excel, txt, images, scanned copies, structured data, and web pages.
Template-based chunking that is described as intelligent and explainable, with multiple template options.
Grounded citations with visualization of text chunking for human intervention and traceable references to reduce hallucinations.
Configurable LLMs and embedding models, with multiple recall paired with fused re-ranking.
Agentic workflow support, MCP, and pre-built agent templates (for example deep research, text2sql, trip planner, and stock market research).
A Python/JavaScript code executor component for agents, run in a gVisor-based sandbox when enabled.

Tech stack

Go (go 1.26.4) as the primary repository language, using Gin (gin-gonic/gin v1.12.0), GORM (gorm.io/gorm v1.25.7), the CloudWeGo eino framework (cloudwego/eino v0.9.12), and OpenTelemetry.
Python backend components (requires-python >=3.13,<3.14) built on Flask/Quart, peewee, and langgraph (==1.2.0).
Document/full-text and vector storage via Elasticsearch by default (go-elasticsearch/v8, elasticsearch-dsl==8.12.0), with an option to switch to Infinity (infinity-sdk==0.7.2) or OpenSearch.
Supporting services: MySQL, MinIO object storage (minio-go/v7, minio==7.2.4), and Redis/Valkey.
Broad LLM and provider SDK coverage including anthropic (==0.76.0), cohere, mistralai, groq, ollama, zhipuai, dashscope/qianfan, voyageai, google-genai, and litellm (==1.82.5); MCP support via mcp>=1.19.0.
Delivered primarily as Docker images (Docker >= 24.0.0, Docker Compose >= v2.26.1); development uses uv for Python dependency management.
Licensed Apache-2.0.

When to reach for it

You need a self-hosted RAG system over heterogeneous document formats rather than a hosted-only or single-format tool.
Citation traceability and human-inspectable chunking matter for your answers.
You want to combine retrieval with agent workflows, MCP tools, or code-executor steps in one platform.
You want to plug in your own choice of LLM and embedding providers instead of being locked to one.
You prefer deploying via Docker Compose and can meet the resource baseline (CPU >= 4 cores, RAM >= 16 GB, Disk >= 50 GB).

When not to reach for it

You are on ARM64 and want turnkey images; all official Docker images are built for x86, and ARM64 requires building your own (switching to Infinity on Linux/arm64 is also not officially supported).
You want a lightweight embeddable library; RAGFlow is a multi-service system requiring Elasticsearch/Infinity, MySQL, MinIO, and Redis.
Your environment cannot meet the stated minimums or run Docker/Docker Compose.
You need the sandboxed code executor without installing gVisor, which is a listed prerequisite for that feature.

Maturity signal

RAGFlow is actively maintained and widely adopted: about 85,600 stars, an Apache-2.0 license, and a last push in July 2026 for a project created in December 2023. Frequent dated updates through 2026 (multi-channel chat, DeepSeek v4, agent memory, MinerU/Docling parsing) and versioned Docker releases (v0.26.4) indicate ongoing, fast-moving development rather than maintenance mode. The roughly 2,100 open issues are consistent with a large, active user base rather than neglect.

Alternatives

Haystack — reach for it when you want a Python-native RAG/LLM framework you assemble in code rather than a full self-hosted server.
LlamaIndex — reach for it when you primarily need a data framework and indexing library to embed in your own application.
Dify — reach for it when your emphasis is an LLM app/agent orchestration platform with a broad app builder rather than deep document parsing.
Infinity — reach for it directly when you only need the vector/full-text database layer; RAGFlow can use it as its document engine.

Notes

RAGFlow can swap its document engine from Elasticsearch to Infinity, a sibling project from the same organization (infiniflow/infinity).
The repository carries agent-oriented tooling for its own development, including .agents/, AGENTS.md, CLAUDE.md, and .github/copilot-instructions.md, plus a documented OpenClaw skill for accessing RAGFlow datasets.
Despite Go being the primary language, a large Python surface remains; the pyproject pins many dependencies with explicit comments explaining CVE mitigations and broken-release avoidance (for example pinning litellm to 1.82.5 and constraining pyasn1, trio, and urllib3).

Wiki: infiniflow/ragflow