Most enterprise RAG pipelines begin in a similar manner: A text parser converts web pages and documents to plain text so that they can be fragmented and indexed for retrieval. That conversion step destroys retrieval cues – and according to new research, it’s responsible for the majority of incorrect answers.
A research team from UC Berkeley, Princeton University, EPFL, and Databricks published a paper this week introducing PixelRAG, a system that skips that conversion entirely. Instead of parsing pages into text, PixelRAG renders them as screenshots, indexing those images and feeding the retrieved tiles directly into a vision-language model reader. Tested on 30 million screenshot tiles covering the entire Wikipedia, it outperforms text-based RAG in six benchmarks, improving accuracy by 18.1% over the text-based baseline.
According to the research team, parsers are the wrong place to look for improvements.
"Improving parsers is an endless process as each website requires special management," lead author and UC Berkeley doctoral student Yichuan Wang told VentureBeat. "Our goal was to find out whether recent advances in VLM made it possible to overcome that entire problem and create a recovery system that works across 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, chunking, and many other handcrafted steps," Wang said. "Each step introduces potential cascade errors and abstractions that take us further away from the original webpage. We were interested in whether we could eliminate most of that complexity and work directly on the rendered page."
Wang also said that parsing inevitably loses information. Images, visual hierarchy, typography, emphasis (for example, bold text), tables, and layout are either discarded or converted to imperfect textual approximations.
"No matter how good a parser becomes, some information is inherently lost during conversion," He said.
The research identifies three ways in which text-based RAG loses the answer before it reaches the reader. All three were measured on SimpleQA, a standard benchmark of 1,000 factual Wikipedia questions:
- Parser loss (36.6% of failures). HTML-to-text conversion destroys the structured content so completely that no text segment in the corpus contains the answer.
-
Rank loss (55.2% of failures). The answer is present in the corpus, but is overtaken by the keyword-dense infobox, which comes in at rank 1 for 75.9% of questions, pushing the answer paragraph to rank 20 or lower.
-
Reader loss (8.2% of failures). The correct content reaches the reader but the flattened structure leads to misattribution.
How PixelRAG works
Unlike a standard LLM that only reads text, a vision-language model takes text as well as images as input, meaning it can read a rendered web page the same way a human does, complete with layout and structure. "For many structured information extraction tasks, we believe that modern VLMs have an inherent advantage because they can jointly reason over both content and layout, rather than relying on a flat text representation," Wang said.
PixelRAG is built around that principle, replacing the text parsing pipeline with a four-step system that works solely on rendered screenshots.
- rendering. Pages are rendered to a fixed 875-pixel viewport using the browser automation library, Playwrite, and cut into 1024-pixel-tall tiles. Wikipedia’s 7 million articles produce approximately 30 million tiles. Assets are cached locally and presented completely offline.
-
Index. Each tile is encoded as a single 2048-dimensional vector using Qwen3-VL-Embedding-2B and stored in the FAISS approximate nearest-neighbor index. The full index runs about 120 GB in fp16 and supports incremental updates without full re-indexing.
-
Training. The retrieval model is fine-tuned on synthetic contrastive data generated from the datastore, 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 of approximately 40,000 pairs on an H100 is completed in less than three hours.
-
storage. Wikipedia requires 5.6 TB for raw screenshot tiles, but the render-on-demand approach eliminates persistent storage: embed all the tiles, delete the screenshots, and re-render the pages on demand at query time. The vector index requires approximately 120 GB.
Six benchmarks, 10x agent token savings and an unsolved problem
Researchers tested PixelRAG on six benchmarks spanning factual Wikipedia QA, table-based queries, multimodal QA, and live news retrieval. He said it outperformed text-based RAGs on all six, including tasks where questions could be answered from text alone. On SimpleQA it reaches 78.8% accuracy versus 71.6% for the strongest text parser, which increases to 42.5% versus 48.8% on structured table queries. Teams need a model in the Qwen3-VL-4B range or above to see benefits. The smaller models outperform text retrieval by more than 12.5 percentage points.
Agent cost benefits are the strongest near-term case for PixelRAG. In benchmark testing, an AI agent using PixelRAG as its search backend ran on 3.6 million prompt tokens compared to 37.5 million for text retrieval, at 2 to 4 times lower cost than alternatives, including Google, while achieving higher accuracy. Image compression can reduce that token budget by another third.
Visual chunking is the main unsolved problem. Text-based RAG systems have spent years refining the way they divide documents into meaningful retrieval units based on topic, section, or semantic content. PixelRAG currently has no equivalent: it cuts pages according to fixed pixel height, meaning that a table or paragraph can be cut in half the middle tile without awareness of content limitations.
"The text retrieval community has spent years studying fragmentation strategies, while visual retrieval has received little attention," Wang said. "We believe this is an important area for future research."
What does this mean for enterprises
The recovery quality issue PixelRAG addresses reflects the broader market shift already underway. VB Pulse Q1 2026 data of qualified enterprise respondents found intent to adopt hybrid recovery tripling from 10.3% in January to 33.3% in March, the fastest growing strategic position in the dataset. PixelRAG’s own authors point to hybrid deployment as the most practical near-term path – layering visual retrieval on top of existing text systems rather than replacing them.
For teams already running RAG pipelines, the path to those savings is more straightforward than a ground-up rebuild.
"A practical route is to use PixelRAG as an enhancement layer alongside existing text retrieval systems," Wang said. "Hybrid retrieval that combines both text and visual search is straightforward and is likely how many production deployments will develop."
<a href