RAGFlow

RAGFlow is a platform for building knowledge-backed AI apps and agent workflows. It is best suited to teams that need control over document retrieval and model connections; its self-hosting requirements make it a poor fit for resource-constrained setups.

Overview

RAGFlow brings document ingestion, search, model connections and agent orchestration together for teams building AI applications around their own information. Its appeal is breadth: the same platform can process varied sources, retrieve relevant material and connect that context to chats or agents. That scope also means teams should weigh operating requirements and plan limits before committing.

Key features

Ingestion and retrieval

The ingestion pipeline cleans and processes multiple data formats into semantic representations. Document sources include Notion, Google Drive, S3, Microsoft Teams, REST API and Sitemap. Search combines vector retrieval with BM25, custom scoring and re-ranking, giving teams several ways to shape retrieval rather than relying on vector search alone. Source citations are supported, a useful fit for knowledge applications where users need to see supporting material.

Agents and models

Visual workflows combine retrieval, tools and MCP servers for agent orchestration. RAGFlow connects online, local and OpenAI-compatible model providers for knowledge bases, chats, search and agents. External MCP servers can connect over Streamable HTTP or SSE; stdio servers need a gateway or another service that exposes a supported endpoint, so not every tool setup connects directly.

By default, RAGFlow validates MCP server URLs and rejects hosts resolving to loopback, private, link-local, reserved or other non-public addresses. That is a sensible safeguard for tool connections, though it may require adjustment to workflows built around internal hosts.

Integration and deployment

HTTP and Python APIs support integration, and an assistant can be embedded in a third-party webpage with an iframe. Docker deployment provides a browser-accessible service. The documented minimum build requirement—4 CPU cores, 16 GB RAM and 50 GB disk—puts self-hosting beyond lightweight hardware. RAGFlow tests ARM64 platforms but does not maintain ARM Docker images; users on Linux or macOS ARM must build an image themselves.

Pricing

The Free plan costs nothing and allows 5 Apps, 1 team member, 0.1 GB dataset storage and 500 credits per month, but has no API key. It is useful for a small initial evaluation, not for API-led integration or substantial datasets.

Starter costs $29 a month and raises the allowance to 50 Apps, 5 team members, 5 GB storage and 5,000 credits per month; an API key is available. It suits small teams moving beyond a trial, though storage and monthly credits remain finite. Pro costs $129 a month and includes unlimited Apps, 20 team members, 50 GB storage, 20,000 credits per month and an API key. That is the stronger fit for larger teams and heavier use, at a higher monthly cost.

Enterprise has custom pricing and lists BYOC and on-premises deployment, dedicated support and a custom SLA. It is aimed at organizations that need deployment options and defined support arrangements. The listed plans make monthly prices and allowances clear, but do not provide a basis for comparing renewal terms.

Platforms

RAGFlow supports API, Linux, macOS, self-hosted and web use. The ARM image caveat matters for teams seeking a ready-made Docker deployment on ARM systems.

Who it's for

The platform is positioned for enterprise use, with solutions for financial services, legal and compliance, manufacturing and education. It fits teams that want one environment for data ingestion, retrieval and agent workflows, and can work within the chosen deployment's resource demands and plan limits. Those looking for a small, API-enabled free tier should look elsewhere: RAGFlow's free plan has no API key.

Pros and cons

• Pros: Combines multi-format ingestion, hybrid search, scoring and re-ranking, which gives teams more retrieval controls in one platform.
• Pros: Connects online, local and OpenAI-compatible models and supports visual agent workflows with external MCP servers.
• Pros: Offers HTTP and Python APIs, iframe embedding, and both cloud-style and self-hosted deployment options.
• Cons: The Free plan has just 0.1 GB storage and 500 monthly credits, with no API key, limiting its usefulness for integrated projects.
• Cons: Self-hosting requires at least 4 CPU cores, 16 GB RAM and 50 GB disk, while ARM Docker images must be built by the user.
• Cons: MCP stdio servers need an intermediary gateway or supported endpoint, which adds setup for those integrations.

Alternatives

Retrieval-Augmented Generation Tools is a useful starting point for comparing the category. Choose AnythingLLM instead if Android or Windows support matters, or if its self-hosted Docker deployment better matches your preferred setup. Flowise is worth considering for a free trial and a free plan that includes 2 flows and assistants, 100 predictions per month and 5 MB storage. I3K RAG Enterprise offers a free, self-hosted Community plan for Linux x86_64, ARM64 or Windows users.

OpenRAG is a free, self-hosted option with no features held back. Dify offers a free Sandbox plan with 200 message credits, one workspace and one member, making it an alternative for a tightly bounded start. Haystack Enterprise Platform has a free Studio plan with one workspace, one user and 100 pipeline hours, plus a free trial. Ragen is a free self-hosted choice with no licence fees, though users pay for their own infrastructure and models. Google Cloud Agent Evaluation is a paid, usage-priced option for teams seeking agent evaluation rather than an integrated RAG and agent platform.

Verdict

Choose RAGFlow if your team needs a configurable path from document ingestion through retrieval to model-powered agent workflows, and can meet the deployment and plan requirements. Its strongest reason to choose is that this toolchain lives in one platform; its clearest reasons to look elsewhere are the free tier's lack of API access and the hardware burden of self-hosting.