Retrieval-augmented generation has always had an awkward bill to pay: the documents you retrieve are expensive to stuff into a context window, and text compression throws away detail irreversibly — a deleted character cannot be recovered. OpticalRAG, posted on arXiv on November 10 (School of Information, Renmin University of China; arXiv:2611.13675), suggests a different diagnosis. The bottleneck may not be insufficient information density, but the choice to compress inside a one-dimensional text sequence. Render a full document page as an image and a modern visual encoder holds it in roughly 100 visual tokens with little semantic loss — a compression ratio the purely textual route cannot reach.
Three components, about thirty million trainable parameters
The pipeline is render, encode, retrieve, distill. Documents are split into 24,000-character chunks, rendered by a deterministic typesetting engine into 640×640 page images, and passed through a frozen DeepEncoder (the SAM-ViT-B plus CLIP combination taken from the DeepSeek-OCR checkpoint), producing 100 visual tokens of dimension 1024 per page. Each visual token is stored individually in an offline FAISS index and retrieved by dot product with a fixed top-K of 256.
Only two pieces are trained. Encoder-Bridged Retrieval projects the LLM's last-layer query states into the visual feature space through a projection shared with the compressor, so no contrastive training is required. Query-Driven Distillation is a lightweight cross-attention compressor that condenses the 256 retrieved visual tokens into 16 using 16 learned queries, after which a VisualToLLMBridge projects them back into the LLM's embedding space.
The DeepEncoder and Qwen2.5-7B-Instruct stay frozen. Only the compressor (14,187,520 parameters) and the bridge (16,529,408 parameters) are trained, 30,716,928 in total. Training runs in two stages: text reconstruction pretraining on 3,800 Wikipedia and C4 samples at a learning rate of 1×10⁻⁴, then task-aware fine-tuning on 11,400 examples from SQuAD v2, DROP, and CoQA at 3×10⁻⁶. The repository ships inference code and the compressor checkpoint, but not the training workflow.
Of 49.68 points, 13.89 come from the multiplier
The headline metric is CAP (compression-aware accuracy): the task score multiplied by a budget-dependent factor, m(B) = ((12000+1)/(B+1))^0.05. Recomputing all 21 cells of the main table from that formula reproduces the published numbers exactly. But the recomputation surfaces something worth reporting.
On LongBench-v2 the raw score is 35.79, the highest of the seven methods in the paper's table — 4.38 points above standard RAG's 31.41 and 5.17 points above the closed-book backbone's 30.62 — while injected tokens fall from 1622.68 to 16, roughly 101 times fewer. CAP reaches 49.68, also the highest. The catch: 35.79 × 1.3882 = 49.68, so 13.89 points, 28% of the reported CAP, are pure compression credit rather than task accuracy.
BAMBOO makes the structure plainer. The raw score of 61.00 actually trails RECOMP's 75.50 and LLMLingua-2's 73.50; what wins is CAP, 84.67 against 81.80. On LooGLE-v2 the raw 23.54 is genuinely highest, yet xRAG injecting a single token with a raw score of 21.06 still lands a CAP of 32.54, just 0.13 behind the paper's 32.67. The reward for very small token budgets is louder than the raw-score gap.
More tokens can hurt
The ablation table shows compression is not a case of more being better. As injected tokens rise from 1 to 64, overall scores run 31.0 (1 token), 31.6 (2), 31.4 (4), 32.8 (8), 35.8 (16), 30.0 (32), and 34.2 (64). Thirty-two tokens scores below eight, and sixty-four still trails sixteen. The paper's explanation is the query-conditioned intrinsic rank r: the compression budget only needs to match r, and on LongBench-v2 r ≤ 16, so exceeding it adds no information and only makes optimization harder.
Not one team's idea
At least three groups are now working the pixel-for-text route. DeepSeek-OCR reported in October 2025 that a page of text compresses to roughly 100 visual tokens, keeping 97% OCR precision below a 10x compression ratio and about 60% at 20x, and on OmniDocBench surpassing GOT-OCR2.0 (256 tokens per page) using only 100 vision tokens. PixelRAG, from June 2026, retrieves web pages as screenshots directly and does better on text-centric benchmarks like NQ and SimpleQA, improving accuracy by up to 18.1% over text baselines and cutting token cost by up to 3x at lower resolution. RAGOCR, in August, trains a GRPO-based compressor to assign each passage a dynamic resolution, beating naive RAG by over 15% in accuracy with one-eighth the input tokens; its learned compression ratios double as reranking scores that outperform Qwen3-Reranker-4B.
OpticalRAG belongs to that line but pushes the compression far harder, to 16 tokens.
Three limits before you ship this
The paper is explicit: evaluation covers multiple-choice questions only, with open-ended generation unverified; a fixed 16-token prefix budget may not suffice for questions requiring evidence spread across distant document segments; and adaptive prefix budgets are left to future work. The repository provides inference code and checkpoints but no training workflow, so reproducing training means building it yourself.
This is closer to a proof of direction than something to copy into production. But the question it raises is worth taking seriously for anyone doing context compression: we keep asking how to make text smaller, when maybe the better question is whether the information can be carried in a different modality.