Breaking

PixelRAG: Screenshot-Based System Outperforms Text Retrieval With 30M Wikipedia Tiles

September 18, 2026 at 13:31 EDT

Foundation Models
Research & Papers
Software Dev & Coding

A new open-source retrieval system from UC Berkeley researchers is replacing text-based document parsing with screenshot tiles, achieving 18% higher accuracy while cutting AI agent token costs tenfold. PixelRAG, released in May 2026, treats web pages and PDFs as visual snapshots rather than extracting and parsing text—a shift that preserves layout, tables, charts, and spatial relationships that conventional text-based retrieval-augmented generation (RAG) destroys.

AI Retrieval · September 2026

PixelRAG Replaces Text Parsing with Visual Tiles

Berkeley researchers open-source a retrieval system that indexes 30 million Wikipedia screenshot tiles, cutting AI token costs 10× while boosting accuracy by 18%.

18.1%

Accuracy gain vs text RAG

10×

Token cost reduction

30M

Screenshot tiles indexed

97%

Storage savings

SimpleQA Benchmark Accuracy

71.6%

Text-Based RAG

78.8%

PixelRAG

+7.2 percentage points improvement on factual Q&A

AI Agent Token Consumption (1 Query)

37.5M

Text Retrieval

3.6M

PixelRAG

10.4× reduction in prompt tokens per query

How It Works

Render pages at 875px width, slice into 1024px tiles
Encode each tile as 2048-dim vector (Qwen3-VL-Embedding-2B)
Store index in FAISS (~120 GB for 30M tiles)
Render-on-demand: delete screenshots, regenerate when needed

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.