AskChem re-indexes chemistry papers as 2.4M atomic, provenance-carrying claims (each tied to a DOI plus a verbatim quote), so an LLM reader grounded in it cites only resolvable sources: 100% of DOIs resolve on CrossRef vs 88.3% for the unaided reader.
You’ve shipped a research-assistant agent that answers questions like “which catalysts hit >90% selectivity for reaction X, and under what conditions?” Today the agent either calls a paper-search API and gets back ranked PDFs (leaving the model to skim and stitch), or it answers from parametric memory and invents DOIs that look right but don’t resolve. Both paths force you to build a verification layer downstream.
The dominant baseline here is Retrieval-Augmented Generation over paper chunks, or agentic deep-research systems like Edison Scientific’s PaperQA that retrieve whole documents and summarize. AskChem argues the retrieval unit itself is wrong: papers are too coarse when the answer lives in one sentence across 40 papers.
The core move is changing what gets stored in the index. Instead of paper chunks, an LLM extractor reads each paper and emits claims: atomic typed assertions like “catalyst Ni SA-N2-C achieves 98% CO Faradaic efficiency,” each carrying a source DOI, a verbatim quote from the paper, and structured fields (reactants, conditions, measurements). If no contiguous quote exists (common in full-text extraction), the claim carries an evidence_locator pointing to where in the paper it came from. 100% of the 2.4M indexed claims carry this provenance.
On top of the shared claim store, AskChem builds three navigational layers. A faceted taxonomy assigns each claim paths like coupling/cross_coupling/suzuki across views (reaction type, substance, technique, mechanism, time), induced from the corpus and then stabilized through synonym normalization and fuzzy clustering. An evidence graph adds 171K typed directed edges between claims (supports, contradicts, extends, derives_from), extracted by a second LLM pass. A separate living taxonomy places papers under governing scientific principles rather than surface topics.
Retrieval is hybrid: FTS5 over claim text, dense vector search, taxonomy-node recall, and paper-level recall, combined with Reciprocal Rank Fusion. An agent-facing query flow looks roughly like:
subqueries = llm_rewrite(user_question) # 3-4 keyword queries
hits = []
for q in subqueries:
hits += hybrid_search(q) # fts5 + vector + taxonomy + paper
claims = rrf_merge_and_diversify(hits, max=40)
answer = llm_synthesize(user_question, grounded_in=claims)
# each cited claim carries source_doi + verbatim_quote
The system exposes the same claim objects through a web UI, REST, an SDK, and an MCP (Model Context Protocol) server, so an agent gets the identical provenance-bearing records a human sees.
The prevailing approach to grounding an LLM on scientific literature is document-level Retrieval-Augmented Generation: retrieve papers or chunks, hand them to the model, hope it cites correctly. This paper shows the opposite. If you make the atomic claim (with its DOI and verbatim quote) the retrieval unit, citation fabrication largely disappears because the model has nothing to hallucinate. The DOI is already attached to the evidence it’s reading. The load-bearing evidence is the DOI-resolvability gap on AskChem-Bench, not the taxonomy or graph features.
•
The headline finding: on AskChem-Bench, grounding a GPT-5.5 reader in AskChem gives 100% resolvable DOIs, versus 88.3% for the same reader with no retrieval. The unaided reader fabricates plausible-looking citations that fail CrossRef lookup.
•
AskChem also has the highest citation density of five tested systems at 18.1 verified DOIs per answer, and the best mean relevance score.
•
The illustrative case: on one CO2-reduction question, the unaided GPT-5.5 fabricates 6 of 14 cited DOIs, while the AskChem-grounded reader has all 22 of 22 resolve.
•
A domain expert audited a stratified sample of 148 evidence-graph edges. After excluding 2 undecidable cases, 143 of 146 had the correct relation type: 97.9% edge-type precision.
•
Honest limitation the paper flags: Edison Scientific’s closed agentic system produces more quantitative detail per answer and a slightly higher on-topic rate. AskChem’s advantage is being open-data, claim-level, interactive-fast, and directly usable as an agent tool, while eliminating DOI hallucination.
Reach for this when you’re building a chemistry research agent and the current pain is that citations don’t resolve, or that answers require aggregating conditions across many papers (e.g., “what Faradaic efficiencies have been reported for CO2-to-CO across all reported electrocatalysts?”). Instead of retrieving PDFs and reparsing them per query, you hit the REST endpoint or MCP (Model Context Protocol) server, get back claims with DOIs and quotes already attached, and feed those to your synthesizer. Your prompt no longer has to say “cite carefully.” There’s nothing to invent.
The system is live at askchem.org with 2.4M claims from 147K papers (1925–2026), a web UI, REST API, SDK, and MCP server. API docs are at askchem.org/api/docs and the AskChem-Bench artifacts (30 cross-paper questions across condition aggregation, temporal tracking, and contradiction surfacing) are at askchem.org/api/benchmark. Extraction uses GPT-5-mini for abstracts and Gemini 3.1 Pro for full-text PDFs; prompts and schemas are released with the source code. The public index exposes claim text and provenance but does not redistribute paywalled full text.
When you need a model to cite real sources, don’t retrieve documents and ask it to cite. Retrieve pre-extracted claims that already carry their citation, and let the model only compose. The fabrication problem then reduces to an extraction-quality problem, which you can audit once and reuse forever.
•
Groundedness is not correctness. 100% of claims are traceable to a source quote, but the paper explicitly notes this does not prove the LLM extractor interpreted each source correctly. A confidently-cited wrong claim is still wrong, just auditable.
•
The benchmark is small: 30 questions, judged by Gemini 3.1 Pro (calibrated to expert labels at 93% agreement). The DOI-resolvability win is robust, but relevance and citation-density claims rest on a thin evaluation.
•
Taxonomy placement and the retrieval contribution of taxonomy-based recall have not been isolated or expert-validated. If the taxonomy is doing less work than the hybrid text+vector search, most of the architecture around the claim store is decorative for citation grounding.