RAGFlow is an Apache-2.0 open-source RAG engine from InfiniFlow that turns documents into a cited, agent-usable knowledge base through layout-aware parsing, template-based chunking, and an agentic retrieval workflow, deployable self-hosted over Docker or through its cloud.
RAGFlow is the largest bet that the retrieval engine itself is the product, and its 92.2k stars measure how badly DIY RAG teams wanted a deployable stack with inspectable chunks and grounded citations, but a disruptive 1.0 rewrite and an unpatched-RCE disclosure are the costs of adopting it.
What it is #
A full engine, not a library: DeepDoc handles layout analysis, OCR, and table recognition, chunking is template-based per document kind (Word, slides, Excel, images, scanned copies, structured data, web pages, and more), and every answer ships with traceable citations plus a chunk view you can correct by hand. Since August 2026 it compiles datasets into agent-usable knowledge artifacts (wikis, graphs, trees, page indexes, mind maps, timelines, skills) and runs agentic retrieval, whose v1.0 release candidates replaced the Low, Medium, High, and Ultra thinking levels with a single Agentic chat mode. Surfaces: a web UI, REST API and Python SDK, an optional MCP server and sandbox executor, Docker deployment, and a hosted cloud. Made by InfiniFlow, with v1.0.0-rc1 (2026-09-29) rewriting the service layer in Go: API, Admin, Ingestor, and Syncer as one Go service, NATS for messaging, Kvrocks for cache and checkpoints.
Status #
Very active and very large: 92,158 stars and 10,979 forks as of 2026-10-13, pushed 2026-10-26 (GitHub API, as of 2026-10-26), with 3.9M Docker pulls (ragflow-stats badge, re-verified 2026-10-16).
The latest release is v1.0.0-rc2 (2026-10-16), the second release candidate of the comprehensive Go rewrite: it keeps rc1 data compatible for direct upgrades, and the one-way migration warning now applies to upgrades from the pre-1.0 v0.27.2 line. The discussion footprint is front-loaded: a 230-point launch thread in April 2024 with 53 comments, 294 comment mentions on Hacker News since the start of 2025 (as of 2026-10-06), and only 2 of those in 2026 itself, a quieting worth weighing against the star count.
Strengths #
- Template-based chunking with a human-in-the-loop chunk view makes the parsing story inspectable, which is the claim most DIY RAG stacks cannot make about their own pipeline.
- Grounded citations and chunk-level traceability ship in the box, the feature teams most often build badly themselves.
- The agent surface is substantial: agentic retrieval modes, knowledge compilation into artifacts agents can read, and an MCP server so coding agents consume the knowledge base directly.
- Scale buys support: 3.9M Docker pulls and a repository this size mean the bug you hit is probably already filed.
Cautions #
- ZeroPath disclosed a post-auth RCE in RAGFlow 0.24 (April 2026) affecting instances that use Infinity for chunk storage, reported to the project on 2026-03-03 as GHSA-vw46-rrp3-c99v and unpatched as of the article, with about 1,918 instances exposed on the public internet; the advisory remains absent from GitHub’s global advisory database (re-checked 2026-10-16).
- The repository’s own security-advisories page carries a published record the global database lacks: eight advisories as of 2026-10-26, including three 2026 criticals with no patched version listed, a Jinja2 template-injection RCE in the Agent Text Processing and Message components (CVE-2026-28797, published 2026-04-01), a Zip Slip RCE in the MinerU parser (CVE-2026-24770, published 2026-01-27), and a Jinja2 template injection in the prompt generator reaching arbitrary OS commands (CVE-2026-45312, published 2026-05-09), plus a 2026 high-severity bypass that lets an authenticated user invoke another user’s private agent through the OpenAI-compatible completions endpoint (GHSA-c6q8-23rp-32cv, published 2026-10-05), alongside 2025’s CodeExec eval RCE, a predictable-token authentication bypass, and a critical SQL injection.
- The 1.0 rewrite is disruptive for pre-1.0 installs by its own release notes: the v0.27.2 data upgrade is irreversible, the local sandbox is dropped, and DeepDoc is CPU-only; rc2 restored Team/Me sharing and keeps rc1 data compatible, so the break lands between v0.27.x and the 1.0 line.
- Self-hosting is a stack, not a binary: the README recommends 4 CPU cores, 16 GB RAM, and 50 GB disk, and the compose file carries MySQL plus a document engine (Elasticsearch, with kernel tuning, or Infinity) alongside the new NATS and Kvrocks services.
- The cloud free tier is deliberately tiny (0.1 GB storage, 500 credits a month, no API key), and the cloud app itself is a login-walled client-side application, so the tiers are only readable on the marketing page.
Pricing #
Self-hosted is free under Apache-2.0. Cloud, as of 2026-10-16 (re-verified unchanged from 2026-10-13): Free at $0 a month (5 apps, 1 member, 0.1 GB storage, 500 credits/month, no API key), Starter at $29 a month (shown struck through from $59; 50 apps, 5 members, 5 GB, 5,000 credits, API key), Pro at $129 a month (struck through from $259; unlimited apps, 20 members, 50 GB, 20,000 credits), Enterprise custom with BYOC, on-premises, and custom SLA.
Price history #
| Date | Plan | Change | Source |
|---|---|---|---|
| 2026-10-06 | Cloud | Baseline: Free $0/mo (500 credits, 0.1 GB, no API key), Starter $29/mo (marked down from $59), Pro $129/mo (marked down from $259), Enterprise custom; self-hosted free (Apache-2.0). |
Compared to #
- Docling: the parser layer RAGFlow bundles as DeepDoc; if you already run a parser and a vector store, RAGFlow replaces that whole pipeline rather than slotting into it.
- Knowhere: the other parse-then-serve bet; Knowhere is hosted per-page parsing with structure-preserving chunks over MCP, RAGFlow is the full engine you deploy.
- LlamaIndex: the framework where you assemble this yourself; RAGFlow sells the assembled, cited, UI-bearing result.
Bottom line #
Recommended for teams that want a deployable, citation-grounded knowledge base for agents and will operate the stack it needs, starting on their hardest documents exactly because the parsing claims are inspectable. Not for embedding search inside an application process (that is Orama’s slot) or for code repositories (that is open-codebase-index’s slot). My disagreeable claim: what RAGFlow actually sells is trust, and the star count says the market’s scarcest RAG ingredient was never layout accuracy, it was being able to see and fix what the retriever read.
Changes #
- 2026-10-06 - Created in the daily refresh’s retrieval entrant scan.
- 2026-10-07 - Added the infiniflow/ragflow star history chart to the Status section.
- 2026-10-16 - Rotation re-verification re-fetched every cited URL live (cloud tiers and Docker pulls unchanged, v1.0.0-rc1 still latest, the ZeroPath advisory still absent from the global database), and the Cautions section gained the repository’s own published advisory record: six advisories including two 2026 critical RCEs (CVE-2026-28797 template injection, CVE-2026-24770 Zip Slip) each listing no patched version.
- 2026-10-17 - Recorded v1.0.0-rc2 (2026-10-16): a single Agentic chat mode replacing the Low/Medium/High/Ultra thinking levels, Team/Me sharing support, and rc1 data compatibility, with the repository activity refreshed (pushed 2026-10-17); the Cautions line now dates the one-way migration break to the v0.27.x-to-1.0 upgrade.
- 2026-10-22 - Recorded the advisory record growing from six to eight: a third 2026 critical with no patched version (CVE-2026-45312, prompt-generator template injection reaching OS commands) and a 2026 high-severity private-agent access bypass (GHSA-c6q8-23rp-32cv), with the repository activity refreshed (pushed 2026-10-22) and v1.0.0-rc2 still latest.
- 2026-10-24 - Refreshed the repository activity (pushed 2026-10-23) and re-checked the advisory record unchanged at eight and v1.0.0-rc2 still latest.
- 2026-10-25 - Refreshed the repository activity (pushed 2026-10-25); v1.0.0-rc2 still latest.
- 2026-10-26 - Refreshed the repository activity (pushed 2026-10-26); v1.0.0-rc2 still latest and the advisory record unchanged at eight.
See also #
- Docling - the open parser whose layout-and-table niche RAGFlow internalizes as DeepDoc
- Knowhere - the hosted counterpart selling document parsing as a per-page service
- Chonkie - the chunking-library approach to the splitting step RAGFlow templates
- AnythingLLM - the assistant-runtime neighbor where the document-chat application pattern lives
References #
https://api.github.com/repos/infiniflow/ragflow - 92,158 stars and 10,979 forks (as of 2026-10-13), Apache-2.0, Go, created 2023-12-12, pushed 2026-10-26, 1,356 open issues (GitHub API, as of 2026-10-26)
https://raw.githubusercontent.com/infiniflow/ragflow/main/README.md - DeepDoc, template-based chunking, knowledge compilation, agentic retrieval modes, Go-native architecture, deployment requirements (re-verified unchanged 2026-10-16)
https://ragflow.io/ - the pricing section: Free $0, Starter $29 (from $59), Pro $129 (from $259), Enterprise custom (re-verified unchanged 2026-10-16)
https://zeropath.com/blog/ragflow-rce-unpatched-vulnerability - the April 2026 post-auth RCE disclosure (GHSA-vw46-rrp3-c99v, Shodan-exposed instances), the critical source (re-verified live and unchanged 2026-10-16)
https://hn.algolia.com/api/v1/items/39896923 - the 230-point launch thread, 2024-04-01
https://cloud.ragflow.io/ - the cloud app, HTTP 200 on 2026-10-16 (a client-rendered JS shell; tiers read from the marketing page instead)