[new plugin] search your PDFs by what’s on the page (figures, scans, small details)
Hi all,
At TopK.io we built a Zotero plugin that searches the page images of your PDFs, not just the text. So you can search for things you only remember seeing.
Some queries that worked on my library (~8k pages):
“mirror selfie at the gym” found a small thumbnail inside a social-feed screenshot in a recommender-systems paper. No text on that page describes it.
“image of an eye” found photos in an ophthalmology paper, and also small eye icons inside diagrams.
“loss curve with a sudden spike” picked the one training plot that has a spike.
“dog tax frm curry” (typo included) found a caption that only exists as pixels in a screenshot.
It also works on scans and drawings where the text isn’t selectable.
How it works: each PDF page is rendered as an image and indexed with TopK’s late-interaction retrieval. Every word in your query gets matched against the parts of the page image, which is why it catches small details.
Please note before trying it:
It isn’t local. Page images are uploaded to TopK, so it’s not a good fit for sensitive documents.
You need a TopK API key. New accounts get $10 of credit. Indexing my 8k pages cost about $2, and storage is about $0.20/month.
Disclosure: I work at TopK and helped build the embedding model.
There’s a demo video in the README: https://github.com/matospiso/zotero-topk
I’d like to know if it’s useful to you, and if you have odd queries or documents you want me to try, post them here.
At TopK.io we built a Zotero plugin that searches the page images of your PDFs, not just the text. So you can search for things you only remember seeing.
Some queries that worked on my library (~8k pages):
“mirror selfie at the gym” found a small thumbnail inside a social-feed screenshot in a recommender-systems paper. No text on that page describes it.
“image of an eye” found photos in an ophthalmology paper, and also small eye icons inside diagrams.
“loss curve with a sudden spike” picked the one training plot that has a spike.
“dog tax frm curry” (typo included) found a caption that only exists as pixels in a screenshot.
It also works on scans and drawings where the text isn’t selectable.
How it works: each PDF page is rendered as an image and indexed with TopK’s late-interaction retrieval. Every word in your query gets matched against the parts of the page image, which is why it catches small details.
Please note before trying it:
It isn’t local. Page images are uploaded to TopK, so it’s not a good fit for sensitive documents.
You need a TopK API key. New accounts get $10 of credit. Indexing my 8k pages cost about $2, and storage is about $0.20/month.
Disclosure: I work at TopK and helped build the embedding model.
There’s a demo video in the README: https://github.com/matospiso/zotero-topk
I’d like to know if it’s useful to you, and if you have odd queries or documents you want me to try, post them here.
Upgrade Storage