PDF-based pipelines are fundamentally lossy and compute-heavy—whether they rely on OCR, GROBID, or LLM-based parsing. They're simply not good enough for accurate, scientific agents at scale.
To fix this, I'm launching ScienceStack API: a lossless, node-based API for scientific papers with LaTeX source, starting with arXiv.
It currently covers 150k+ arXiv papers, mainly in CS, Math, and Physics.
I’m giving away 5× 3-month Pro keys to early commenters who are building in this space (scientific tooling, agents, copilots, RAG etc). I’d love to hear what you’re working on
It works! After the initial data load (big paper), the scrolling and performance works nicely.
Can visit at sciencestack.ai/arxiv/2105.10386
Note: no support for nomenclature/index yet.
I'm also working on refactoring the data/json to a streaming model (right now it's one big json dump on load)
Yea there were several attempts (including ar5iv), and distill.pub is no longer active + Semantic Scholar is PDF-based.
None quite made the full use of HTML or have a robust conversion system. Jeff Dean's post is awesome - though using Gemini 3 is compute intensive and may still hallucinate in the end (I'm using a source-based latex to json parser). And the output is still...not very interactive.
To fix this, I'm launching ScienceStack API: a lossless, node-based API for scientific papers with LaTeX source, starting with arXiv.
It currently covers 150k+ arXiv papers, mainly in CS, Math, and Physics.
Every paper also ships with a WYSIWYG interactive reader at sciencestack.ai/paper/{arxivId}. Example: https://www.sciencestack.ai/paper/2512.24601v1
I’m giving away 5× 3-month Pro keys to early commenters who are building in this space (scientific tooling, agents, copilots, RAG etc). I’d love to hear what you’re working on