Scientific Papers as APIs,
not documents
We parse scientific papers so you don't have to.
Don't waste compute on PDFs — focus compute on science.
Parsed from LaTeX source. No OCR. No hallucinations. 100ms latency.
Query papers like code
Paper: 1706.03762v7
Click any node<a id="sec-1"></a>
## 1: Introduction
---
Recurrent neural networks, long short-term memory [[hochreiter1997]](#bib-hochreiter1997) and gated recurrent [[gruEval14]](#bib-gruEval14) neural networks in particular, have been firmly established as state of the art approaches in sequence modeling and transduction problems such as language modeling and machine translation [[sutskever14]](#bib-sutskever14), [[bahdanau2014neural]](#bib-bahdanau2014neural), [[cho2014learning]](#bib-cho2014learning). Numerous efforts have since continued to push the boundaries of recurrent language models and encoder-decoder architectures [[wu2016google]](#bib-wu2016google), [[luong2015effective]](#bib-luong2015effective), [[jozefowicz2016exploring]](#bib-jozefowicz2016exploring).
Recurrent models typically factor computation along the symbol positions of the input and output sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden states $h_t$ , as
...Every node contains full content from LaTeX source.
Available as Markdown, LaTeX, and JSON.
The core problem
PDFs weren't built for machines
PDF extraction gives you strings and glyphs. Instead, we parse LaTeX to give you a semantic graph with stable IDs, relationships, and metadata.
PyMuPDF / GROBID / Nougat / LLMsWe call our particular attention "Scaled Dot-Product Attention" (Figure 2). The input consists of queries and keys of dimension dk, and values of dimension dv. Attention(Q, K, V ) = softmax( QKT √dk )V The two most commonly used attention functions are additive attention [2], and dot-product attention. Dot-product attention is identical to our algorithm, except for the scaling factor.
✗ No way to find "equation 1"
✗ No link to Figure 2
✗ Can't extract just this section
✗ No parent/child relationships
GET /api/v1/papers/1706.03762v7/nodes?nodeId=sec:3.2.1&format=markdown## 3.2.1: Scaled Dot-Product Attention
We call our particular attention "Scaled Dot-Product Attention" (Figure [fig:2](#fig-2)). The input consists of queries and keys of dimension $d_k$, and values of dimension $d_v$.
$$
\mathrm{Attention}(Q,K,V) = \mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$
The two most commonly used attention functions are additive attention [\[bahdanau2014\]](#bib-bahdanau2014), and dot-product attention. Dot-product attention is identical to our algorithm, except for the scaling factor.✓ Same section, any format you need
✓ Query any node type (equation, figure, table, etc.)
✓ Figures have CDN URLs
✓ Stable nodeIds and parent/child relationships
We preserve structure & math
PDF extraction is lossy, LaTeX is not
We parse every paper directly from LaTeX source — no OCR, no prediction, no hallucination.
Attention(Q, K, V ) = softmax( QKT √dk )V\mathrm{Attention}(Q,K,V) = \mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})VRaw glyph extraction. No structure, no LaTeX — just Unicode characters.
<formula>Attention(Q,K,V) = softmax(QK^T/sqrt(d_k))V</formula>\mathrm{Attention}(Q,K,V) = \mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})VStructured XML, but math is not reliably recoverable as LaTeX.
\nabla_\theta \mathcal{L} = \frac{1}{N}\sum_{i=1}^{N} \frac{\partial \ell_i}{\partial \theta_j}\nabla_\theta \mathcal{L} = \frac{1}{N}\sum_{i=1}^{N} \frac{\partial \ell_i}{\partial \theta}Hallucinated subscript: θ_j vs θ. Predicted, non-deterministic, high latency — and costs tokens per call.
Try it yourself
Preview responses from "Attention Is All You Need" (1706.03762v7).
Endpoints available in Markdown, LaTeX, and JSON.
Any section, equation, figure, or table — in your preferred format
#### 3.2.1: Scaled Dot-Product Attention
We call our particular attention "Scaled Dot-Product Attention" (Figure [fig:multi-head-att](#fig:multi-head-att)). The input consists of queries and keys of dimension $d_k$, and values of dimension $d_v$.
$$
\mathrm{Attention}(Q, K, V) = \mathrm{softmax}(\frac{QK^T}{\sqrt{d_k}})V \tag{1}
$$
The two most commonly used attention functions are additive attention [\[bahdanau2014neural\]](#bib-bahdanau2014neural), and dot-product (multiplicative) attention...API at a glance
Every endpoint you need to build with scientific papers.
| Endpoint | Free |
|---|---|
/searchFind papers by topic, author, arXiv ID | FREE |
/papersBrowse papers by category or field e.g. "machine-learning", "cs.CV" | FREE |
/papers/{id}/overviewTOC, figure/equation/table refs, AI summaries | FREE |
/papers/{id}/figuresImage URLs for vision models | quota |
/papers/{id}/nodes?types=equationFilter by type: equation, table, math_env, algorithm | quota |
/papers/{id}/nodes?nodeId=sec:3.2.1Access any node with stable IDs | quota |
/papers/{id}/contentFull paper as Markdown, LaTeX, or text | quota |
/papers/{id}/referencesBibliography with Semantic Scholar enrichment | quota |
Once you access a paper, you get unlimited requests to that paper for the rest of the month. Learn more
Optimized for scientific agent workflows
Browse free, deep dive once, use unlimited. Ideal for fully autonomous agents.
Search
Find papers by topic or arXiv ID
Overview
Get TOC, figures, equations, AI summaries via /overview
Deep Dive
Access full content per paper (uses 1 paper quota)
Unlimited
Fetch that paper's equations, figures, nodes for the rest of the month
How do I know the data is good?
Every paper in our API comes with an interactive reader , as proof that our parsing works.
The reader is a live demonstration of our API data. Hover any citation, equation, or figure — that's the same structured data you get via API.
- Hover citations & equations for instant previews
- Dependency graphs showing how concepts connect
- Annotations that sync across devices
- Export to PDF, LaTeX, Markdown, or JSON
- Dark mode & mobile-friendly
Build Tools and Copilots with ScienceStack
Query any section, equation, figure, or citation — as Markdown/LaTeX/JSON, full metadata, and parent/child relationships.
150k+
Papers indexed
<100ms
Avg response time
99.9%
Uptime
v1
Stable API
Scientific Copilots
Build AI tools that understand papers like researchers do — cite specific equations, reference exact figures.
"Explain equation 3 from the attention paper"
Citation-Aware RAG
Ground every answer in verifiable sources. Stable node IDs (eq:3, fig:2) enable precise attribution.
Link answers directly to paper sections
Bulk Paper Analysis
Extract all equations from 100 transformer papers in minutes. Compare methods systematically.
No PDFs. No OCR. Just data.
Knowledge Graphs
Build citation networks from structured bibliographies. References linked to arXiv IDs and DOIs.
Map how papers connect
Used by researchers and developers at
Simple, usage-based pricing
Start free. Pay only when you need more. /search and /overview are always free.
Free
Try it out
Explore the API. Search papers and access structured content.
Start Free- Unlimited /search queries
- /overview for any paper
- Markdown, LaTeX, text export
- 10 papers/month deep access
Pro
For researchers & developers
AI summaries, per section summaries, and node-level AST.
- Everything in Free
- AI summaries (paper + section)
- Node-scoped AST export
- 200 papers/month, $0.10 overage
Team
For scientific tools & pipelines
Full paper AST for RAG and agent workflows.
- Everything in Pro
- Full paper AST export
- Higher rate limits
- 1,000 papers/month, $0.05 overage
Ready to build?
Get your API key and create amazing tools with ScienceStack.