RAG-Anything

RAG-Anything: All-in-One RAG Framework — Zirui Guo, Xubin Ren, Lingrui Xu, Jiahao Zhang, and chao-huang (hkuds; arXiv:2510.12323, 14 Oct 2025). MIT-licensed, 22.6k stars, 426 commits.

The cluster’s first multimodal system. Every RAG method this corpus holds — GraphRAG, lightrag, hipporag2, raptor, leanrag, memgraphrag — indexes text and only text. RAG-Anything’s premise is that this is a mismatch with the documents people actually have: “modern knowledge repositories are inherently multimodal, containing rich combinations of textual content, visual elements, structured tables, and mathematical expressions.”

Built on lightrag, by the same group and the same senior author, and it inherits that system’s incremental-update property rather than restating it.

The pipeline

Five stages, and the first one is the part no text-only system needs:

1. Document parsing — MinerU by default (OCR, table extraction, GPU acceleration), with Docling for Office/HTML and PaddleOCR as alternatives. Handles PDF, DOC/DOCX/PPT/PPTX/XLS/XLSX, images, and plain text. LibreOffice is a hard dependency for Office formats.
2. Multimodal content understanding — separate concurrent pipelines per modality, with document hierarchy preserved rather than flattened.
3. Analysis engine — dedicated analyzers for images, tables, LaTeX equations, plus an extension point for custom types.
4. Knowledge graph construction — multimodal elements become first-class entities, not captions attached to text. Cross-modal relationships are inferred automatically, and document structure survives as explicit belongs_to chains.
5. Modality-aware retrieval — vector similarity fused with graph traversal, ranked with awareness of content type.

The design claim

The framing the paper leads with is that multimodal content should be treated as interconnected knowledge entities rather than isolated data types, implemented as dual-graph construction — one structure carrying cross-modal relationships, one carrying textual semantics, unified in a single representation. Retrieval is correspondingly hybrid: structural navigation plus semantic matching, so evidence for one answer can span modalities.

That is a recognizable move within the cluster. lightrag fuses graph and vector retrieval for text; this fuses them across modalities, with the modality boundary handled at index time rather than at query time.

What the evidence actually shows — read this before citing it

The abstract asserts “superior performance on challenging multimodal benchmarks,” “significant improvements over state-of-the-art methods,” and gains that “become particularly pronounced on long documents.” It names no dataset, no baseline, and no number.

The paper itself does better: the evaluation uses DocBench and MMLongBench, against lightrag, GraphRAG, GNN-RAG, HippoRAG, RAPTOR, ArchRAG, MM-GraphRAG, VideoRAG and VisRAG. The numeric tables could not be extracted from the PDF on ingest (compressed content stream), so this page records the benchmarks and baselines as verified and the figures as unread — not as absent. Anyone leaning on a specific margin should open the paper.

Two things are settled regardless. The evaluation is author-run, like lightrag‘s and unlike graphrag-bench‘s. And the comparison set is the cluster’s own leaderboard, which graphrag-bench already showed to be task-conditional — graph methods beat vanilla RAG on multi-hop and summarization and lose on simple fact lookup, at one to two orders of magnitude more tokens. A multimodal extension of a graph method inherits that shape unless it shows otherwise.

Why it matters here

It opens an axis the cluster did not have. The corpus’s evidence thread has been about how much graph structure is worth and at what token cost; every entry in it assumed the source documents were text. RAG-Anything’s parsing stage is the admission that in practice they are not — and it puts the hardest part of the problem, extraction fidelity from a PDF, outside the retrieval design entirely, in MinerU. The cluster’s cost and accuracy arguments say nothing about that stage yet.

Related

lightrag · chao-huang · hkuds · knowledge-graph · graphrag-bench · from-local-to-global-graphrag · hipporag2 · raptor · leanrag · memgraphrag · retrieval-augmented-generation · entity-resolution