Academic Formula Parser

Search arXiv for academic papers on any topic and extract the mathematical formulas and equations from each paper — pulling the real LaTeX out of arXiv's HTML5 full-text build when one exists, and falling back to scanning the abstract otherwise.

No API key, login, or credentials are required. The actor uses only arXiv's free, public Atom query API and the public arxiv.org/html/<id> full-text pages.

---

What it does

For a given search query the actor:

1. Calls the public arXiv query API (export.arxiv.org/api/query) and retrieves up to maxResults papers, sorted by relevance.
2. For each paper, extracts the title, authors, abstract, arXiv ID, publication date and abstract URL.
3. Extracts equations:
   • If Parse Full Text is enabled (default), it fetches the paper's HTML5 build at arxiv.org/html/<arxivId> and pulls the genuine LaTeX out of the MathML <annotation encoding="application/x-tex"> blocks that arXiv embeds for every rendered equation. Older papers often have no HTML build — in that case the actor falls back to the abstract.
   • It also scans the title + abstract free text for inline/display LaTeX ($...$, $$...$$, \[...\], \(...\), \begin{equation}...), symbolic relations in prose (e.g. E = mc^2, dS/dt >= 0), and standalone math tokens (subscripts/superscripts, Greek letters, scientific notation).
4. De-duplicates the detected expressions (HTML LaTeX takes priority over abstract heuristics) and pushes one record per paper to the dataset.

The formulaSource field on every record tells you whether the equations came from the paper's full text (html-fulltext) or only the abstract (abstract).

---

Input

Field	Type	Default	Description
query	string	"thermodynamics"	Topic or keyword to search arXiv for. Matched against all fields (all:<query>).
maxResults	integer	10	Number of arXiv papers to fetch and parse. Clamped to the range 1–100.
parseFullText	boolean	true	When true, fetch each paper's HTML full text to extract real LaTeX equations. When false, only the abstract is scanned (faster, one request per run instead of one per paper).

Example input

{
  "query": "Schwarzschild black hole entropy",
  "maxResults": 5,
  "parseFullText": true
}

---

Output

One dataset record is pushed per paper. Example (truncated):

{
  "query": "attention is all you need transformer",
  "title": "Energy-Gated Attention and Wavelet Positional Encoding",
  "author": "Jane Doe, John Smith",
  "authors": ["Jane Doe", "John Smith"],
  "abstract": "We introduce ...",
  "equations": ["\\leq 6", "\\omega_{i}", "\\sigma_{i}"],
  "equationDetails": [
    { "expr": "\\leq 6", "type": "latex-html" },
    { "expr": "\\omega_{i}", "type": "latex-html" }
  ],
  "equationCount": 83,
  "formulaSource": "html-fulltext",
  "url": "http://arxiv.org/abs/2510.12345v1",
  "arxivId": "2510.12345v1",
  "published": "2025-10-15T12:00:00Z",
  "source": "arXiv"
}

Output fields

Field	Type	Description
query	string	The search query that produced this record.
title	string	Paper title.
author	string	Authors joined into a single comma-separated string ("Unknown" if none listed).
authors	string[]	Authors as an array.
abstract	string	Paper abstract.
equations	string[]	De-duplicated list of extracted formula/equation strings (LaTeX or symbolic).
equationDetails	object[]	Same expressions with a type tag: latex-html, latex-display, latex-inline, latex-env, relation, or expression.
equationCount	integer	Number of equations in equations.
formulaSource	string	html-fulltext if equations came from the paper's HTML full text, otherwise abstract.
url	string	arXiv abstract page URL.
arxivId	string	Bare arXiv identifier (e.g. 2510.12345v1 or hep-th/9901001).
published	string	Publication timestamp (ISO 8601).
source	string	Always "arXiv".

If the query returns no papers, a single marker record is pushed with equationCount: 0 and a note field explaining that arXiv returned no results.

---

Authentication & setup

None. The actor requires no API keys, tokens, or accounts. It uses only arXiv's public endpoints.

---

Limitations

• arXiv only. This actor searches arXiv exclusively. It does not query other databases (PubMed, IEEE, Semantic Scholar, etc.).
• Full-text LaTeX requires an HTML build. arXiv only generates HTML5 (with embedded LaTeX) for many — not all — papers, mostly more recent ones. Papers without an HTML build fall back to abstract-only extraction, which usually yields few or zero equations because abstracts rarely contain typeset math. This is expected; check formulaSource to see which path was used.
• Heuristic abstract extraction. Formula detection in plain abstract prose is regex/heuristic based, so it can miss unusual notation or occasionally include a near-miss expression. The html-fulltext path (real MathML LaTeX) is the high-fidelity source.
• Equations are returned as raw LaTeX/symbol strings. They are not rendered to images or evaluated. No solving, simplification, or symbolic computation is performed.
• Rate limits. With parseFullText enabled the actor makes one extra HTTP request per paper to arxiv.org. Large maxResults values mean more requests; arXiv may throttle very aggressive use.

---

How equations are typed

type	Meaning
latex-html	Real LaTeX extracted from the paper's HTML5 MathML annotations (highest fidelity).
latex-display	Display-math LaTeX found in text ($$...$$, \[...\]).
latex-inline	Inline LaTeX found in text ($...$, \(...\)).
latex-env	LaTeX environment (\begin{equation}...\end{equation}, align, etc.).
relation	A symbolic relation detected in prose (e.g. E = mc^2, dS/dt >= 0).
expression	A standalone math token (subscript/superscript, Greek letter, scientific notation).