StarTrail-org/pixelrag-bench

PixelRAG reader benchmark

Question answering over retrieved Wikipedia screenshots. Every question comes with the top 5 screenshot tiles that the PixelRAG retriever returned for it from a 26.3M-tile index of English Wikipedia, so a vision-language model can be evaluated as the reader of a visual RAG pipeline without hosting the index or the 4 TB tile corpus.

The question sets, retriever and grader are those of the LoRA-retriever column of Table 1 in the paper.

Configs

config	source	questions	query image
nq	Natural Questions validation, first 1000 questions with at least 2 annotators agreeing	1000	no
nq_tables	NQ-Tables dev (one question appears twice in the source)	1068	no
simpleqa	SimpleQA, first 1000 questions whose sources include a Wikipedia page	1000	no
mmsearch	MMSearch, end2end	300	171 of 300
encyclopedic_vqa_landmarks	Encyclopedic-VQA test, Google Landmarks, automatic questions	749	yes

Each config has a single test split.

from datasets import load_dataset
ds = load_dataset("StarTrail-org/pixelrag-bench", "simpleqa", split="test")
row = ds[0]
row["question"], row["answer"]
row["retrieved_images"][:3]   # PIL images, best first

Fields

field	content
id	example id from the source dataset
question	the question
instructions	answer-format instruction the paper appended to the prompt (empty for some configs)
answer	reference answer as the grader receives it
query_image	image that is part of the question, or null
retrieved_images	5 tiles, ranked by retrieval score; each is a slice of a rendered page, 875 px wide and at most 1024 px tall
retrieved_scores	retrieval scores for the 5 tiles
retrieved_urls	Wikipedia article slug of each tile
original_data	JSON: the full source example and the retrieval settings, as consumed by the grader

Paper protocol

The paper's reader sees the top 3 tiles. The system prompt, for questions without a query image:

You are a research assistant who answers questions based on provided evidence.
Use <think></think> tags to show your reasoning if needed.
Answer the question directly and concisely based ONLY on the provided evidence.

and the user turn is the question, then the 3 tiles, then instructions.

For questions with a query image:

You are a research assistant who answers questions based on retrieved visual evidence.
You will receive: (1) a text question, (2) a query image, and (3) retrieved Wikipedia evidence images.
Use the query image and evidence images to answer the question.
Use <think></think> tags to show your reasoning if needed.
Answer the question directly and concisely.

and the user turn is one text block (Question: <question>, a blank line, The first image is the query image. The following images are retrieved Wikipedia evidence. Answer the question based on the evidence., then instructions after a blank line), followed by the query image and the 3 tiles.

The code that builds these messages is build_messages in eval/lib/llm.py.

Scores reported in the paper for the Qwen3.5-4B reader with the top 3 tiles (thinking enabled for MMSearch only): NQ 58.7, NQ-Tables 48.8, SimpleQA 78.8, MMSearch 28.3, Encyclopedic-VQA landmarks 45.1.

Scoring

The grader is eval/lib/grader.py in the PixelRAG repository. For simpleqa it applies the SimpleQA grading template and for mmsearch and encyclopedic_vqa the WorldVQA judge prompt, with gpt-4.1-2025-04-14 as the judge. For nq and nq_tables it defaults to exact match; --llm-judge switches to the GPT-4.1 judge, which is what the paper's numbers use. It reads a JSONL file with one line per question:

import json
with open("preds.jsonl", "w") as f:
    for row, pred in zip(ds, predictions):
        od = json.loads(row["original_data"])["original_data"]
        f.write(json.dumps({"example_id": row["id"], "original_data": od, "final_response": pred}) + "\n")

git clone https://github.com/StarTrail-org/PixelRAG && cd PixelRAG/eval && uv sync --frozen
OPENAI_API_KEY=... PYTHONPATH=. .venv/bin/python -m lib.grader simpleqa preds.jsonl
# nq and nq_tables: add --llm-judge to match the paper

Grader task names: nq, nq_tables, simpleqa, mmsearch, encyclopedic_vqa.

Construction

Retrieval ran against search_index_lora_vit_ckpt200_v2 from StarTrail-org/pixelrag-faiss-indexes (Qwen3-VL-Embedding-2B with the PixelRAG LoRA, IVF with 4096 lists, nprobe 128, 2000 for SimpleQA), with the query instruction Retrieve images or text relevant to the user's query.. Questions with a query image were retrieved with text and image together. The export scripts are eval/dump_bench.sh and eval/pack_bench.py in the PixelRAG repository, added in PR #165.

6 of the 20,585 retrieved tiles, in 5 questions, no longer exist in any copy of the corpus. Those entries keep their rank, score and URL, and their image is null.

License

The tiles are renders of English Wikipedia and are CC BY-SA 4.0. Questions, answers and query images keep the licenses of their source datasets: Natural Questions, NQ-Tables, SimpleQA, MMSearch, Encyclopedic-VQA and, for its query photos, the Google Landmarks Dataset v2.

Citation

@misc{wang2026pixelragwebscreenshotsbeat,
      title={PIXELRAG: Web Screenshots Beat Text for Retrieval-Augmented Generation},
      author={Yichuan Wang and Zhifei Li and Zirui Wang and Paul Teiletche and Lesheng Jin and Matei Zaharia and Joseph E. Gonzalez and Sewon Min},
      year={2026},
      eprint={2606.28344},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2606.28344},
}