Scholar Feed

Retrieval-augmented generation

VisRAG

VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documents

Superseded baseline#57 of 1,179 most-superseded · first seen Oct 14, 2024

Superseded — cited as a baseline and beaten by newer methods

3 papers critique it · 1 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites VisRAG as a baseline.

However, current retrieval-augmented pipelines for multimodal document QA are inherently static, introducing two key limitations.
— MARA: A Multimodal Adaptive Retrieval-Augmented Framework for Document Question Answering
“Early visual RAG systems, such as ColPali faysse2024colpali and VisRAG yu2024visrag, pioneered the use of document page snapshots as retrieval units, directly feeding images to VLMs to bypass error-prone OCR pipelines and preserve crucial layout semantics. However, this page-level granularity creates a fundamental bottleneck: retrieving entire pages introduces substantial irrelevant visual content, which dilutes the generator's attention and forces high-resolution pages into limited visual token budgets, often sacrificing detail and increasing hallucination risk tanaka2025vdocrag, wang2025vidorag.”
— AgenticOCR: Parsing Only What You Need for Efficient Retrieval-Augmented Generation
“VisRAG preserves document layouts as images but misses granular relationships”
— RAG-Anything: All-in-One RAG Framework

Beaten on benchmarks

Head-to-head results where a newer method reports beating VisRAG. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.