On July 30, 2026, researchers from New York University and Matterstack posted the AskChem paper on arXiv. It describes claim-centered retrieval: the system indexes single claims, not papers or chunks. AskChem holds 2.4M claims from 147K chemistry papers, and each claim carries its source DOI and a verbatim quote.
What changed: the retrieval unit moves from the paper to a typed claim with its own DOI and quote. Who's affected: developers building RAG over scientific papers, and agents that need citations they can check. What to do: test claims where traceability matters, and keep chunks where extraction cost dominates.
The authors are Bing Yan, Gregory Wolfe, Stefano Martiniani and Kyunghyun Cho. All benchmark figures below come from the authors' own benchmark, not an independent test.
Chunks cannot cite a single proposition
A chunk points to a location in a paper, so a model citing it cannot show which sentence backs which statement. Say you ask whether a catalyst works in water. Chunk RAG returns the third paragraph of a methods section. A claim store returns the one statement that the catalyst was tested in water, its result, and the DOI and quote behind it. That gap matters because models invent references. The OpenScholar study in Nature reports that GPT-4o "fabricated citations in 78–90% of cases" when asked to cite recent literature.
AskChem's own appendix shows the same failure without retrieval: "6 of 14 cited DOIs do not resolve on CrossRef" in one example answer.
Left, chunks overlap and lose their link to a statement. Right, each claim card keeps its own provenance and links to related claims.
Each claim carries its DOI and a quote
Every paper becomes atomic, typed claims, each grounded by a source DOI and a verbatim quote or an evidence locator. Two extractors do the work. A high-throughput one reads abstracts at scale. A deeper one reads full-text PDFs to catch hypotheses, limitations and surprising findings that abstracts often omit.
Claims, taxonomy and graph share one SQLite store
The storage layer is modest: SQLite with FTS5 full-text search and a vector index, served through FastAPI. No separate vector database or graph database is needed: claims, taxonomy and graph live in one SQLite file, so a small team can reproduce the serving layer. The expensive step is extraction, an LLM pass over every paper before anything is indexed. Over the claims sit two navigation structures:
- Taxonomy: 307K populated nodes for browsing and filtering claims.
- Evidence graph: 171,342 typed edges between claims. A domain-expert author checked a stratified sample of 148 edges, and 143 of 146 decidable edges had the right type (97.9%).
Citations nearly doubled, and DOIs rose 11.7 points
On AskChem-Bench (30 cross-paper chemistry questions on condition aggregation, temporal tracking and contradiction surfacing, GPT-5.5 reader), grounding in AskChem lifted resolvable DOIs from 88.3% to 100%. That is 11.7 points over a strong baseline. The bigger gain is citation density: 18.1 per answer against 9.6 without retrieval, about 1.9 times as many.
| System | DOI existence (%) | Citations per answer | Grounded specificity | Relevance (0–3) | On-topic ≥ 2 (%) |
|---|---|---|---|---|---|
| GPT-5.5 alone | 88.3 | 9.6 | 8.1 | 1.66 | 65.8 |
| GPT-5.5 + AskChem | 100 | 18.1 | 5.9 | 2.15 | 86.6 |
| Paperclip | 100 | 7.5 | 0.5 | 1.72 | 57.8 |
| Edison Scientific | 99.1 | 10.7 | 29.2 | 2.07 | 89.7 |
| NotebookLM | 93.7 | 7.9 | 0.1 | 1.84 | 78.9 |
Grounded specificity is the paper's count of quantitative tokens that share a sentence with a citation marker, so a higher score means more numbers tied directly to a citation. Read the table per system before drawing a winner. AskChem is not alone at 100% DOIs, because Paperclip ties it. Edison Scientific's PaperQA-family agent leads grounded specificity, 29.2 against AskChem's 5.9, and edges it on on-topic answers. AskChem leads citation density and relevance.
Agents connect through REST, SDK or MCP
AskChem offers a web interface plus REST, SDK and MCP access for AI agents. An MCP client can therefore query the claim store directly. If you already wire agents to MCP servers, this is one more tool to point at.
Choose claims when traceability beats extraction cost
Use claims when a wrong citation is expensive and answers span many papers. Keep chunks when you cannot afford an LLM pass over every paper before indexing. A quick checklist:
- Do answers combine findings across papers? Claims help.
- Must every statement link to a DOI and quote? Claims help.
- Do you need dense numeric detail? Check Edison-style agents too.
- Is your corpus outside chemistry? Expect to design your own claim types.
For off-the-shelf options, compare literature review tools. For pipeline operations, see running LLM pipelines in production. Confidential papers may call for local AI for sensitive papers.
What is still unconfirmed
- Correctness: the authors note that provenance checks establish traceability, not full semantic correctness: a claim can cite a real DOI and quote and still be misread during extraction. No claim-extraction error rate is reported, according to an alphaXiv summary.
- Domain: the claim schema includes chemistry-specific fields, so porting to another field means redesigning those claim fields and the extraction, not just re-running the pipeline.
- Scale of the test: 30 questions and one reader model. No latency or cost figures.
- Public access: we could not confirm open-source code, a public dataset or a Hugging Face ranking. Treat those as unconfirmed.
- Breadth: OpenScholar indexes 45 million papers and 236 million passage embeddings at passage level, a different trade-off.
