Most enterprise RAG pipelines start the same way: a text analyzer converts web pages and documents into plain text so they can be broken down and indexed for search. This conversion step destroys search signals and is responsible for the majority of false positives, according to a new study.
A research team from UC Berkeley, Princeton University, EPFL, and Databricks published a paper this week. PixelRAG, a system that skips this conversion entirely. Instead of parsing pages in text, PixelRAG renders them as screenshots, indexes those images, and passes the resulting tiles directly to a visual language model reader. Tested on 30 million screenshot tiles covering the entire Wikipedia, it outperforms text-based RAG on six criteria, improving accuracy by up to 18.1% compared to text-based benchmarks.
Analysts are the wrong place to look for a fix, according to the research team.
"Developer improvement is a never-ending process, as each website requires special handling," Lead author and UC Berkeley PhD student Yichuan Wang told VentureBeat. "Our goal was to investigate whether recent advances in VLMs allow us to bypass the entire problem and learn how to build a search engine that works on websites without site-specific engineering."
HTML parsers destroy the search signals that enterprise RAG depends on
The goal of the researchers was to create a clean end-to-end architecture.
"Modern web RAG pipelines often involve rendering, parsing, cleaning, parsing, and many other manual steps," Wang said. "Each stage introduces potential cascading errors and abstractions that take us away from the original web page. We wondered if we could remove much of this complexity and work directly on the page being rendered."
Wang also noted that analysis inevitably loses information. Images, visual hierarchy, typography, emphasis (eg, bold text), tables, and layout are either removed or reduced to imperfect text approximations.
"No matter how good the parser is, some information is fundamentally lost during the conversion," he said.
The study identifies three ways in which text-based RAG loses the answer before it reaches the reader. All three were measured against SimpleQA’s standard benchmark of 1,000 actual Wikipedia questions:
-
Parser loss (36.6% of hits). HTML-to-text conversion destroys the structured content so completely that no piece of text in the corpus contains the answer.
-
Rating loss (55.2% of success). The answer is present in the corpus, but for 75.9% of queries, it is dominated by dense infoboxes with the rank 1 keyword, pushing the answer paragraphs to 20 or less.
-
Reader loss (8.2% of hits). The right content reaches the reader, but a flat structure leads to incorrect distribution.
How PixelRAG works
Unlike standard LLM, which only reads text, the visual language model accepts images as input alongside text, meaning it can read a displayed web page like a human, with the layout and structure intact. "For many structured data mining tasks, we believe that modern VLMs have a unique advantage because they can reason together about both content and layout rather than relying on a flattened textual representation." Wang said.
PixelRAG is built on this principle, replacing the text parsing pipeline with a four-stage system that works entirely on rendered screenshots.
-
Rendering. Pages are rendered in a fixed 875-pixel viewport using Playwright, a browser automation library, and sliced into 1024-pixel-high tiles. Wikipedia’s 7 million articles produce about 30 million tiles. 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 at the FAISS-estimated nearest-neighbor index. The full index runs up to about 120GB on fp16 and supports incremental updates without full re-indexing.
-
Training. The search model was adapted to synthetic contrast 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 core of the language model and the visual encoder. Training on approximately 40,000 pairs is completed in less than three hours on an H100.
-
Storage. Raw screenshot tiles for Wikipedia require 5.6 TB, but the rendering-on-demand approach eliminates persistent storage: deploy all tiles, delete screenshots, and re-render the requested pages on request. Vector index requires about 120 GB.
Six benchmarks, 10x agent token savings, and one unresolved issue
The researchers tested PixelRAG on six benchmarks, including actual Wikipedia QA, table-based queries, multimodal QA, and live news search. They said it outperformed text-based RAG on all six issues, including tasks where questions could only be answered from text. In SimpleQA, it reaches 71.6% vs. 78.8% for the most powerful text parser, and expands to 42.5% vs. 48.8% in structured table queries. Teams need Qwen3-VL-4B or higher class models to see the benefits. The smaller models outperform text search by more than 12.5 percentage points.
Agent cost advantage is the strongest near-term case for PixelRAG. In the benchmark test, the AI agent using PixelRAG as search support worked with 3.6 million operational tokens against 37.5 million for text search, which achieved 2-4 times lower cost than alternatives, including Google, while achieving higher accuracy. Image compression can reduce this token budget by another third.
Visual fragmentation is a major unsolved problem. Text-based RAG systems have spent years dividing documents into meaningful search units based on topic, section, or semantic content. PixelRAG currently has no equivalent: it slices pages to a fixed pixel height, meaning a table or paragraph can be sliced into half-tiles without knowing the content boundaries.
"The text search community has been studying fragmentation strategies for years, while visual search has received less attention." Wang said. "We think this is an important area for future research."
What this means for businesses
The search quality issue for PixelRAG addresses reflects a broader market shift already underway. VB Pulse Q1 2026 found that qualified enterprise respondents intend to triple hybrid search from 10.3% in January to 33.3% in March, the fastest-growing strategic position in the database. PixelRAG’s own authors point to hybrid positioning as the most practical near-term path – layering visual search on top of existing text systems rather than replacing them.
For teams already running RAG pipelines, the path to this savings is simpler than a major rebuild.
"A practical way is to use PixelRAG as an enhancement layer alongside existing text search engines." Wang said. "Hybrid search, which combines both text and visual search, is simple and likely to evolve as many production deployments."
