Chunk document pages as horizontal image bands: visual-context retrieval with a dense index
tomaarsen · x · 2026-11-03
Perplexity's pplx-embed-v2-context leverages its vision-language backbone to chunk document pages as images: split the page's visual tokens into horizontal bands, encode them together, and keep one vector per band.
This lets you search parts of a page with a regular dense vector index while retaining surrounding visual context — useful for complex layouts where text extraction loses structure. The paper's figure walks through the pipeline from visual patches to band-aligned token sequences with SEP separators.
More from Research
- Nature: Connectome simulations identify a three-neuron circuit generating fly walking rhythms — patrickmineault · 2026-11-03
- AI researcher: models editing weights is real but the self-improvement hype overclaims — paul_cal · 2026-11-03
- D&A Turns Any VLM into a Zero-Shot Robot Reward Model, SOTA on 3 Benchmarks — DJiafei · 2026-11-03
- CT-Perturb: LLMs Reach 82% of Full-Data Cell Perturbation Prediction with 10% of Measurements — david_van_dijk · 2026-11-03
- Christian Kroer's Book 'Games, Markets, and Online Learning' Now in Print, Free PDF Available — bremen79 · 2026-11-03
- Open-source ML curriculum on GitHub covers math foundations to ML engineering — Charming-Ad-4323 · 2026-11-03