infiniflow/ragflow

RAGFlow is worth a serious look if your RAG problems start at the document layer — scanned PDFs, borderless tables, mixed office formats — rather than at the vector-store layer. Budget real setup time and expect to validate parsing quality on your own corpus, not the demo.

The problem

You want to ask questions over a pile of internal documents and get answers with citations. The usual route is to extract text, chunk it, embed it and wire up a chat UI; the failure mode is that extraction mangles the parts that matter most — tables, multi-column layouts, scanned pages — and every downstream answer inherits those errors. RAGFlow moves the work to a visual, template-driven ingestion pipeline: documents are parsed with layout awareness, chunks are shown for human review, and retrieval runs against that cleaned context. The result is an answer you can trace back to a chunk you have actually looked at, instead of a black-box similarity hit.

What's different

The mechanism is deep document understanding plus explicit chunking templates, and the README leans on that combination as the reason for "quality in, quality out" (project overview). Chunking is visual and editable, so you can inspect and correct segmentation before it becomes retrieval truth, and answers carry traceable citations. Above that, the project has been layering agentic retrieval: the model can decompose a hard question, retrieve in several rounds and check evidence, with Low/Medium/High/Ultra thinking modes controlling how much work happens per query. Knowledge Compilation, added in August 2026, further turns document and dataset content into structured artifacts like wikis, graphs, mind maps, timelines and "skills" (release notes).

The trade-off is surface area. Each new layer — agents, compilation templates, thinking modes — is another thing to configure and another source of behavior to debug, and the more autonomy you grant retrieval, the harder it is to reason about why a particular chunk was or was not used. Treat the agentic modes as an option to evaluate against plain retrieval on your own questions, not as a default improvement.

Who should try it

It fits teams whose bottleneck is heterogeneous source material: finance or legal documents, scanned reports, spreadsheets, slide decks, or internal wikis with tables that ordinary text extraction flattens. If you need cited answers over such a corpus and are willing to review chunking, the visual pipeline is the reason to pick this over assembling parsers yourself. It fits less well if you already have a clean text corpus and a working retrieval stack — the ingestion machinery buys you little, and the extra configuration is pure overhead. A hosted option exists at RAGFlow Cloud for those who want to see the pipeline before committing to local deployment.

Requirements and limitations

The project provides two documented paths: the cloud service, and a local deployment that you have to operate yourself. Local deployment is a multi-service stack — the repository's primary language is Go but the system spans a broader set of components — so expect to manage services, storage and model configuration rather than just run a single binary. The agentic layers depend on a capable model; thinking modes and context compilation add model calls and processing time, so question complexity directly affects response latency and cost.

Parsing quality is also corpus-specific. An early commenter praised the layout recognizer's handling of one-column tables that a paid API missed, but noted there was no benchmark and the result could have been luck (HN comment, 2024-04-01). That caveat still frames the right evaluation: test on your own worst documents, not on a sample PDF.

HotRepo's take

HotRepo's tracked history is thin: four daily-board observations from August 2026, the latest showing 88,511 stars on 2026-08-15, and a weekly Go-board rank of #17 on 2026-10-27. That is not enough to judge trend persistence. The specific thing to verify before adopting is how the current ingestion pipeline handles your hardest table or scan — the README's later releases and the 2024 parsing reports are several versions apart, and only your own documents settle the question.