Multi-Vector Visual Document Indexes Can Be Inverted to Reveal Page Content

A new study shows that ColPali-style multi-vector visual document retrievers can be inverted: storing about a thousand patch vectors per page, they can leak up to 47% of page words and allow retrieval back to source pages 98.4% of the time, raising privacy concerns for third-party vector databases.

2026-10-08 ~ 2026-10-08 · 2 related posts

• Multi-vector visual document indices can be inverted: 47% of words recovered, 98.4% source-page recall — Zhuchenyang Liu · 2026-10-08
• ColPali-style visual document indices can be inverted: 47% of words recovered, source page ranked first 98.4% — _reachsumit · 2026-10-08