Most enterprise RAG pipelines start the same way: a text parser converts web pages and documents into plain text so they can be fragmented and indexed for retrieval. That conversion step destroys retrieval signals and, according to new research, is responsible for most incorrect responses.
A research team from UC Berkeley, Princeton University, EPFL and Databricks published a paper this week presenting PixelRAG, a system that skips that conversion entirely. Instead of parsing pages into text, PixelRAG renders them as screenshots, indexes those images, and sends the retrieved tiles directly to a visual language model reader. Tested on 30 million screenshot tiles covering all of Wikipedia, it outperforms text-based RAG on six benchmarks, improving accuracy by up to 18.1% over text-based baselines.
According to the research team, analyzers are the wrong place to look for solutions.
"Improving parsers is a never-ending process because each website requires special handling," Yichuan Wang, lead author and UC Berkeley doctoral student, told VentureBeat. "Our goal was to explore whether recent advances in VLM make it possible to avoid that whole problem and build a recovery system that works on all websites without site-specific engineering."
HTML parsers destroy the recovery signals that enterprise RAG depends on
The researchers’ goal was to develop a clean end-to-end architecture.
"Modern web RAG pipelines often involve rendering, parsing, cleaning, sharding, and many other craft stages." Wang said. "Each stage introduces potential cascading errors and abstractions that take us further away from the original web page. We were interested in whether we could remove most of that complexity and operate directly on the rendered page."
Wang also pointed out that when analyzing, information is inevitably lost. Images, visual hierarchy, typography, emphasis (e.g., bold text), tables, and layout are discarded or become imperfect textual approximations.
"No matter how good a parser is, fundamentally some information is lost during the conversion," said.
The research identifies three ways in which text-based RAG misses the answer before reaching the reader. All three were measured on SimpleQA, a standard benchmark of 1000 objective questions from Wikipedia:
-
Loss of the analyzer (36.6% of failures). Converting HTML to text destroys the structured content so completely that no text fragment in the corpus contains the answer.
-
Loss of range (55.2% of failures). The answer exists in the corpus, but is outnumbered by information boxes full of keywords that reach rank 1 for 75.9% of the queries, leading paragraphs containing answers to rank 20 or lower.
-
Loss of readers (8.2% of failures). The right content reaches the reader, but the flattened structure causes misattribution.
How PixelRAG works
Unlike a standard LLM that only reads text, a visual language model takes images as input along with text, meaning it can read a rendered web page like a human does, with the layout and structure intact. "For many structured information extraction tasks, we believe that modern VLMs have an inherent advantage because they can jointly reason about both content and layout rather than relying on a flattened text representation." Wang said.
PixelRAG builds on that principle, replacing the text analysis process with a four-stage system that operates entirely on rendered screenshots.
-
Representation. Pages are rendered using Playwright, a browser automation library, in a fixed 875 pixel viewport and divided into 1024 pixel high tiles. Wikipedia’s 7 million articles produce approximately 30 million mosaics. Assets are cached locally and displayed completely offline.
-
Indexing. Each tile is encoded as a single 2048-dimensional vector using Qwen3-VL-Embedding-2B and stored in a FAISS approximate nearest neighbor index. The full index takes up approximately 120 GB on fp16 and supports incremental updates without the need to completely re-index.
-
Training. The recovery model is fitted from contrasting synthetic data generated from the data warehouse, using dynamic hard negative mining to filter out false negatives. LoRA, a lightweight fine-tuning method that updates a small fraction of the model weights, is applied to both the language model backbone and the visual encoder. Training with approximately 40,000 pairs is completed in less than three hours with a single H100.
-
Storage. Raw screenshot tiles for Wikipedia require 5.6 TB, but an on-demand rendering approach eliminates persistent storage: embed all tiles, delete screenshots, and re-render pages on demand at query time. The vector index requires approximately 120 GB.
Six Benchmarks, 10x Agent Token Savings, and One Unsolved Problem
The researchers tested PixelRAG on six benchmarks covering Wikipedia QA, table-based queries, multimodal QA, and live news retrieval. They said it outperformed text-based RAG on all six, even on tasks where questions can be answered with text alone. In SimpleQA it achieves an accuracy of 78.8% compared to 71.6% for the most powerful text analyzer, expanding to 48.8% compared to 42.5% in structured table queries. Teams need Qwen3-VL-4B class models or higher to see the benefit. The smallest models lag behind text retrieval by more than 12.5 percentage points.
The agent cost advantage is the strongest near-term argument for PixelRAG. In benchmark tests, an AI agent using PixelRAG as a search engine executed 3.6 million prompt tokens versus 37.5 million for text retrieval, at a cost 2 to 4 times less than alternatives, including Google, while achieving higher accuracy. Image compression can reduce that symbolic budget by a further third.
Visual fragmentation is the main unsolved problem. Text-based RAG systems have spent years perfecting how to divide documents into meaningful retrieval units based on topic, section, or semantic content. PixelRAG currently has no equivalent: it splits pages based on a fixed pixel height, meaning a table or paragraph can be cut in half without regard to content boundaries.
"The text retrieval community has spent years studying chunking strategies, while visual retrieval has received much less attention." Wang said. "We believe this is an important area for future research."
What this means for businesses
The recovery quality issue PixelRAG addresses reflects a broader market shift that is already underway. VB Pulse Q1 2026 Data from qualified business respondents found that intent to adopt hybrid recovery tripled from 10.3% in January to 33.3% in March, the fastest-growing vantage point in the data set. PixelRAG’s own authors point to hybrid implementation as the most practical path in the short term: overlaying visual retrieval on top of existing text systems rather than replacing them.
For teams already running RAG pipelines, the path to those savings is simpler than a rebuild from scratch.
"A practical path is to use PixelRAG as an enhancement layer alongside existing text retrieval systems." Wang said. "Hybrid retrieval that combines visual and text search is straightforward and is likely to evolve in many production deployments."
