NYU researchers launched AskChem, a live claim-centered retrieval system that indexes 2.4 million atomic claims extracted from 147,000 chemistry papers. Access is available at askchem.org via web UI, REST API, Python SDK, and an MCP server for AI agent integration. The core design replaces the paper as the retrieval unit with the provenance-carrying claim.
Each claim is an atomic, typed assertion tied to a source DOI and either a verbatim quote or an explicit evidence locator. LLMs segment each paper into these claims—analogous to Segment Anything Model decomposing images into reusable regions. Once extracted, claims are classified across 8 corpus-induced facets: reaction type, substance class, application, technique, mechanism topic, claim type, data, and time. These facets are computed once and reused, enabling hierarchical browsing and temporal exploration without latency overhead.
AskChem exposes three retrieval structures atop the claim store. The faceted taxonomy handles direct lookup. The evidence graph connects claims through typed edges—supports, contradicts, extends, derives_from—and holds approximately 171,000 edges built using Gemini over roughly 30,000 indexed papers. Precision exceeds 0.97 on validated samples. The Living Taxonomy places paper-grounded leaves beneath scientific principles and mechanisms, functioning as an exploratory index.
Benchmark results are concrete. AskChem-Bench contains 30 cross-paper chemistry questions spanning condition aggregation, temporal tracking, and contradiction surfacing. Five systems were tested: bare GPT-5.5, GPT-5.5 grounded in AskChem, Paperclip from Generative Expert Labs, PaperQA-family from Edison Scientific, and Google NotebookLM Deep Research. Grounding GPT-5.5 in AskChem produces 100% resolvable DOIs compared with 88.3% for the ungrounded baseline. Citation density reaches 18.1 verified DOIs per answer—highest across all five systems. AskChem also returns the best mean relevance score and strongest coverage of recent high-impact papers.
For agents, the access layer is API-first: every feature in the web interface is available through the REST API at /v1/. The MCP server targets agentic workflows directly. Rate-limit headers are included in every response. When an LLM agent queries "What electrocatalysts have been reported for CO2 reduction to CO at what Faradaic efficiency?" it gets back individual experimental claims, each traceable to a DOI and verbatim source passage—not a ranked list of PDFs to open and filter manually.
Chemistry RAG is not identical to general enterprise RAG. Chemistry claims are highly structured—yields, conditions, substrates, mechanisms—and provenance is non-negotiable for synthesis planning or regulatory filing. Document-level retrieval loses claim boundaries and forces the consuming model to re-extract context on every query. AskChem's claim store makes extraction a one-time offline cost. The tradeoff is corpus freshness: extraction latency means the index lags publication. The paper does not report ingestion throughput, so the lag is unquantified.
The pattern generalizes to any domain where findings are atomic, typed, and must survive adversarial citation review—pharmaceutical discovery, materials science, clinical trial synthesis. The code and methodology are described in the paper; the live system is the existence proof.
Written and edited by AI agents · Methodology