RAGFlow is the heavyweight option in self-hosted document AI: an Apache 2.0 RAG engine built around layout-aware document parsing, with agent workflows bolted on since mid-2026. It sits at 91.7k GitHub stars as of October 2026, up from 139-star obscurity when it first hit trending lists. The question this article answers is not “is RAGFlow good” — it is whether its parsing quality justifies a 16 GB RAM footprint when AnythingLLM does 80% of the job on 2 GB.

Short verdict: if your documents are scanned PDFs, multi-column reports, or anything with tables, RAGFlow is the only one of the three that parses them properly. If your documents are clean text and Markdown, it is overkill, and the lighter tools win on setup time and hardware cost.

What license does RAGFlow use, and is it really open source?

RAGFlow is Apache 2.0 — verified from the LICENSE file at github.com/infiniflow/ragflow, October 2026. There is no open-core license trap on the engine itself: commercial self-hosting is unrestricted, no BSL or SSPL clauses, no regional carve-outs. InfiniFlow (the company behind it) sells a managed cloud and enterprise tier, but the self-hosted Docker deployment is the full product.

The current stable release is v0.27.1, with v1.0.0-rc1 published September 29, 2026 — the project is in the final stretch toward a 1.0 that moves the backend to a unified Go service. For production self-hosting today, stay on v0.27.1; release candidates have shipped breaking config changes in this project before.

For comparison: AnythingLLM is MIT (v1.15.0, June 25, 2026), and Flowise is Apache 2.0. All three are genuinely FOSS — this comparison has no license loser, which is rarer than it should be in 2026.

What does RAGFlow actually do differently?

RAGFlow’s differentiator is DeepDoc, its built-in document understanding layer: layout analysis, OCR, and table recognition running on ONNX Runtime before any text ever reaches the embedding model. Most RAG tools — AnythingLLM included — extract raw text and chunk it by character count. That works for clean text and falls apart on real-world PDFs: multi-column layouts get read in the wrong order, tables turn into word soup, and scanned pages extract as nothing at all.

DeepDoc instead classifies regions (heading, paragraph, table, figure), OCRs scanned content, and applies template-based chunking that matches the document type — a resume, a paper, and a legal contract get chunked differently. You can visually inspect the chunks it produced and correct them, which no other tool in this comparison offers.

Since v0.25.1 (April 30, 2026) RAGFlow has also grown agent capabilities: a visual workflow builder with multi-agent pipelines, MCP integration, sandboxed code execution, and data source connectors for Confluence, Google Drive, Notion, S3, GitHub, and Slack. The agent builder is real, not marketing — but it overlaps heavily with what Flowise does with a two-year head start. Buy RAGFlow for the parsing; treat the agent layer as a bonus.

It connects to local LLMs via Ollama or Xinference and to any OpenAI-compatible endpoint, so a fully air-gapped deployment works.

How do you self-host RAGFlow?

The supported path is Docker Compose, and the minimum is real: 4 CPU cores, 16 GB RAM, 50 GB disk, Docker >= 24.0.0, Docker Compose >= v2.26.1. No GPU is required — embedding and OCR run on CPU, though a GPU speeds up ingestion of large document sets.

$ git clone https://github.com/infiniflow/ragflow.git
$ cd ragflow/docker
$ docker compose -f docker-compose.yml up -d
# ... pulls complete ...
$ docker logs -f ragflow-server
    ____   ___    ______ ______ __
   / __ \ /   |  / ____// ____// /____  _      __
  / /_/ // /| | / / __ / /_   / // __ \| | /| / /
 / _, _// ___ |/ /_/ // __/  / // /_/ /| |/ |/ /
/_/ |_|/_/  |_|\____//_/    /_/ \____/ |__/|__/
* Running on all addresses (0.0.0.0)

The web UI then answers on port 80. First login prompts you to register a local admin account and point it at a model provider (an Ollama host on your LAN works).

A problem you will actually hit: on most Linux hosts the Elasticsearch container exits a few seconds after startup with exit code 78, and the UI never comes up. The cause is the kernel’s vm.max_map_count default (65530) being below Elasticsearch’s requirement. The fix:

sudo sysctl -w vm.max_map_count=262144

Add vm.max_map_count=262144 to /etc/sysctl.conf to make it survive reboots, then docker compose up -d again. If you switch the doc engine from Elasticsearch to Infinity (InfiniFlow’s own database) in the compose config, this step is unnecessary — but Elasticsearch is the default and the better-trodden path.

RAGFlow vs AnythingLLM vs Flowise: which should you pick?

Pick by what your documents look like and how much machine you have. All three verified October 2026:

	RAGFlow v0.27.1	AnythingLLM v1.15.0	Flowise v3.x
License	Apache 2.0	MIT	Apache 2.0
Minimum RAM	16 GB	2 GB	4 GB
PDF/table parsing	DeepDoc: layout analysis, OCR, table recognition	Basic text extraction	Depends on loader node you wire up
Chunk inspection	Visual, editable per chunk	Limited settings, opaque retrieval	Manual, per-flow
Agent workflows	Visual builder, MCP, code sandbox (since v0.25.1)	Simple agent skills	Core strength: visual LangChain graphs
Local LLM support	Ollama, Xinference, OpenAI-compatible	Ollama + 30 or more providers	Ollama + most providers
Setup time to first answer	30-60 min (multi-container stack)	About 10 min (single container or desktop app)	About 20 min
GitHub stars (Oct 2026)	91.7k	about 65k	about 45k

The decision reduces to three questions. Are your documents scanned, tabular, or multi-column? Only RAGFlow handles those well — the other two feed garbage text to a good retriever. Is your corpus clean text (Markdown, docs, transcripts)? AnythingLLM gets you the same end result in a tenth of the setup time and an eighth of the RAM. Is the RAG part trivial but the pipeline logic complex? Flowise — it is an orchestration canvas first and a document tool second.

Boundary worth stating: none of these is an enterprise document management system. No audit trails, limited SSO (RAGFlow and AnythingLLM gate some of this behind paid tiers), no records retention. That layer is on you.

What does it cost to run RAGFlow in 2026?

RAGFlow’s real cost is the 16 GB RAM floor, and the 2026 DRAM price crisis makes that floor expensive. A 32 GB DDR5 kit runs $399-$479 as of September 2026 (Tom’s Hardware RAM price index) — roughly 4x its mid-2025 price — so “just upgrade the old office PC” is no longer a $60 decision. If an existing machine already has 16 GB or more, RAGFlow costs you electricity: a small always-on x86 box draws 15-40 W idle, roughly $20-$60/year at US average rates (see our always-on AI server electricity breakdown).

On a rented VPS, 16 GB RAM is mid-tier money, not entry-tier — price it against how many documents you actually have before committing to a monthly bill. And remember RAGFlow itself does not need the GPU; if you want fast local generation behind it, the GPU question is the same one covered in runaihome’s hardware guides (https://runaihome.com/blog/ryzen-ai-max-395-strix-halo/ for the current value picks).

When should you NOT use RAGFlow?

Do not deploy RAGFlow if any of these describe you:

• Your corpus is clean text. Markdown, code docs, plain PDFs that copy-paste cleanly — DeepDoc adds latency and RAM for zero retrieval gain. Use AnythingLLM or Open WebUI’s built-in RAG.
• You have less than 16 GB RAM to spare. RAGFlow under-resourced means Elasticsearch OOM-kills and silent ingestion failures. The minimum is the minimum.
• You want a personal “chat with my files” tool. The multi-container stack (Elasticsearch/Infinity, MinIO, Redis, MySQL, server) is team-scale infrastructure. For one user it is maintenance burden with no payoff — our self-hosted maintenance cost analysis applies in full.
• You need the agent builder, not the RAG. Flowise is lighter and more mature for pure workflow orchestration.

What to actually run

Prices as of October 2026, all verified in the sections above:

Your situation	Run this	Cost	Where
Clean-text docs, one user, want results today	AnythingLLM on any 8 GB machine you own	$0 software	Setup guide
Scanned PDFs/tables, team knowledge base, 16 GB box available	RAGFlow v0.27.1 self-hosted	$0 software + $20-$60/yr power	github.com/infiniflow/ragflow
Same, but no spare 16 GB machine	16 GB cloud VPS (non-GPU)	pay monthly	Vultr
Want GPU-fast ingestion/generation without buying a card	Rented RTX 3090, from $0.07/hr	pay per hour	Vast.ai
Upgrading an 8/16 GB desktop to run it at home	32 GB DDR5 kit, $399-$479	$399-$479	Check price

If you are wiring RAGFlow’s retrieval into a coding workflow as a codebase knowledge layer, aicoderscope covers the editor side: https://aicoderscope.com/blog/kilo-code-vs-opencode-vs-cline-2026/.

FAQ

Does RAGFlow require a GPU? No. OCR, layout analysis, and embedding run on CPU via ONNX Runtime; the minimum spec is 4 CPU cores and 16 GB RAM with no GPU. A GPU only matters if you also host the generation model on the same box — then standard Ollama VRAM math applies (about 19 GB for a 32B model at Q4_K_M).

Can RAGFlow run fully offline/air-gapped? Yes, after initial image and model downloads. Point it at a local Ollama or Xinference instance for both chat and embedding models and it makes no external calls. The optional cloud connectors (Google Drive, Notion, Slack) are off unless configured.

Is RAGFlow’s v1.0.0-rc1 safe to run? Not for anything you care about yet. The 1.0 line rewrites the backend as a unified Go service; rc1 landed September 29, 2026. Stay on v0.27.1 until 1.0 stable has a point release behind it.

Sources

• RAGFlow GitHub repository — license, releases, system requirements (accessed Oct 5, 2026)
• RAGFlow v1.0.0-rc1 release — Sep 29, 2026
• AnythingLLM documentation — version and requirements (accessed Oct 5, 2026)
• Flowise GitHub repository — Apache 2.0 license
• Tom’s Hardware RAM price index — DDR5 pricing, Sep 2026

What self-hosting actually costs

Real cost breakdowns for self-hosted AI: hardware floors, power, maintenance hours, and the honest comparison against paying for it. No spam, unsubscribe anytime.

Email address Subscribe