News · 2026-07-31

AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers

A team at New York University led by Bing Yan, with Stefano Martiniani and Kyunghyun Cho, launched AskChem, a chemistry literature system that changes what a search returns. Instead of a ranked list of papers, it retrieves individual claims - 2.4 million of them, extracted from 147,000 papers, each carrying a source identifier and a verbatim quote. Grounding a language model in it produced 100% resolvable citations, against 88.3% without retrieval.

Key facts

• AskChem indexes 2.4 million atomic claims drawn from 147,000 chemistry papers.
• On its own benchmark, grounding a GPT-5.5 reader in AskChem yields 100% resolvable DOIs versus 88.3% without retrieval, and the highest citation density among five systems tested.
• Launched July 30, 2026; live at askchem.org, with code public. It drew 285 upvotes on Hugging Face's daily papers feed, the highest of the day.
• Primary source: AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis.

The problem AskChem targets is specific to how science actually gets read. A chemist trying to establish whether a particular catalyst works under particular conditions is not looking for a paper. They are looking for a sentence, and the sentence is scattered across forty papers that each discuss something adjacent. Existing search systems return the forty papers and leave the extraction, the provenance checking and the cross-paper assembly as manual work - as the authors put it, "scientists and AI agents need to locate relevant information, verify their provenance, and assemble cross-paper answers manually."

AskChem's move is to do that extraction once, up front, and make the result the searchable unit. Each paper is converted into atomic, typed claims. Each claim is grounded by a DOI plus a verbatim quote or an explicit evidence locator - so the assertion arrives with its receipt attached. Over that shared store the system layers several ways in: a faceted taxonomy for hierarchical browsing, an evidence graph linking claims to one another through relations, and an exploratory taxonomy organising indexed papers under scientific principles.

The analogy is the difference between a library catalogue and an index of quotations. A catalogue tells you which books mention a topic. An index of quotations tells you what was actually asserted, by whom, and where to check - which is what you needed in the first place.

The design decision with the longest reach is that the same claim store is exposed to people and to machines identically, through a web interface plus REST, SDK and MCP access. An AI agent doing literature work queries the same objects a human browses. That matters because the alternative - agents scraping ranked PDFs and summarising them - is precisely the pipeline that produces confident, unverifiable paragraphs.

The measured benefit is narrow and real. On the authors' own benchmark, a GPT-5.5 reader grounded in AskChem produced citations that all resolved, against 88.3% unresolvable-free without retrieval, and the highest citation density of five systems compared. In plain terms: the references it gives you exist, and there are more of them.

There is a quieter contribution in the choice to make the taxonomy explicit and revisable rather than fixed. Chemistry does not have one settled hierarchy of concepts, and any system that pretends otherwise will misfile the interesting work - which is usually the work that does not sit anywhere obvious. AskChem separates a stabilised faceted taxonomy used for reliable retrieval from an exploratory living taxonomy that arranges papers under scientific principles, and labels the second as exploratory rather than authoritative. That is a modest-sounding design decision with real consequences for whether a search system helps or quietly narrows what a field can find.

Here is the caveat that deserves as much attention as the launch. Provenance is not truth. A claim carrying a resolvable DOI and a verbatim quote proves that a statement can be traced back to text. It does not prove the extraction interpreted that text correctly, that the underlying paper was methodologically sound, that the finding has been replicated, or that several apparently independent claims are not all restating one flawed upstream result. A study published this month makes the danger concrete: planting a single credible-looking false document in a research agent's retrieval pool flipped more than half of its final reports to the false conclusion, even though a separate verifier could correctly identify the document as misleading. AskChem makes a bad claim vastly easier to locate, audit and retract. Something else still has to decide whether it is true.

Primary source, verified: read the paper → (arXiv 2607.28618)

Ground Truth. (2026, July 31). AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers. Ground Truth. https://groundtruth.day/news/askchem-makes-the-claim-not-the-paper-the-unit-of-search.html

@misc{groundtruth:askchem-makes-the-claim-not-the-paper-the-unit-of-search,
  title  = {AskChem indexes 2.4 million individual chemistry claims instead of 147,000 papers},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/askchem-makes-the-claim-not-the-paper-the-unit-of-search.html}
}

Topics: science · retrieval · agents · chemistry · provenance · tools