Summary (Overview)

• AskChem introduces a claim-centered infrastructure for chemistry literature synthesis, shifting the retrieval unit from documents to atomic, provenance-carrying claims.
• Each claim is an atomic, typed assertion extracted from a paper, grounded by a source DOI and a verbatim quote or evidence locator.
• The system exposes three complementary structures over a shared claim store: a stabilized faceted taxonomy, an evidence graph linking claims via typed relations, and an exploratory living taxonomy organized by scientific principles.
• AskChem currently indexes 2.4M claims from 147K papers (1925–2026) and provides a web interface, REST API, SDK, and MCP server for both human users and AI agents.
• On the AskChem-Bench evaluation suite (30 cross-paper chemistry questions), grounding a GPT-5.5 reader in AskChem yields 100% resolvable DOIs (vs. 88.3% without retrieval) and the highest citation density (18.1 verified DOIs per answer) among five tested systems.

Introduction and Theoretical Foundation

The paper addresses a fundamental mismatch in current literature-search systems: chemists often need to assemble answers from specific findings scattered across many papers (e.g., "What electrocatalysts have been reported for CO₂ reduction to CO, and at what Faradaic efficiency?"), but existing interfaces primarily return ranked lists of documents. This leaves scientists and AI agents to manually locate relevant information, verify provenance, and assemble cross-paper answers.

The key insight is that LLM agents now survey papers and plan experiments (Boiko et al., 2023), but inherit the limitations of document-level retrieval tools. When asked to answer from parametric memory, an LLM may fabricate plausible-looking citations (Agrawal et al., 2024). AskChem addresses this by making the provenance-carrying scientific claim—rather than the paper—the central object for search, browsing, and agent access. The authors draw an analogy to the Segment Anything Model (Kirillov et al., 2023), which decomposed images into reusable regions; AskChem uses LLMs to segment papers into claims.

Methodology

Claim-Centered Representation

The central object is a Claim: an atomic, typed scientific assertion extracted from a paper, grounded by a source DOI and a verbatim quote. Each claim includes structured fields (e.g., reactants, conditions, measurements, materials) and an extraction confidence score. The system stores several structures under the same claim identity:

• Source: paper-level metadata (DOI, venue, year, citation count, authors via OpenAlex)
• TreeNode: places a claim in one or more faceted taxonomy paths
• Edge: records a typed relation between two claims

Claim Extraction Pipelines

AskChem uses two complementary extraction pipelines:

1. High-throughput extractor: processes abstracts at scale (GPT-5-mini)
2. Deep extractor: reads full-text PDFs (Gemini 3.1 Pro), capturing claim types often absent from abstracts (hypotheses, limitations, surprising findings)

Each extraction returns structured JSON validated against the claim schema, including required provenance fields, numeric ranges, and chemistry-specific fields.

Evidence Graph Construction

The system adds a relation-extraction layer that emits typed, directed edges between claims:

• cites_as_evidence, supports, extends, contradicts, derives_from
• Each edge has confidence scores and provenance
• Current graph: 171,342 typed edges with 97.9% edge-type precision (domain-expert audit of 148 edges)

Stabilized Faceted Taxonomy

AskChem induces category paths while digesting papers, then stabilizes them through:

• Canonical top-level routing
• Synonym normalization
• Fuzzy clustering of near-duplicate subcategories

The taxonomy defines multiple operational views: reaction type, substance class, application, technique, mechanism topic, claim type, data (measurements), time, and source authors.

Retrieval Method

Hybrid /search combines:

• FTS5 claim-text retrieval
• Paper-level recall
• Taxonomy-node recall
• Dense-vector recall
• Reciprocal rank fusion (Cormack et al., 2009)

Empirical Validation / Results

Corpus Statistics

• 2.4M grounded claims from 147K source papers (1925–2026)
• 99.9% DOI-verified (CrossRef)
• 66% full-paper claims (1.6M), 34% abstract-only (833K)
• 307K populated taxonomy nodes

Claim-Type Distribution

AskChem-Bench Results (30 cross-paper questions)

Key findings:

• AskChem achieves 100% resolvable DOIs (vs. 88.3% without retrieval)
• Highest citation density (18.1 verified DOIs per answer)
• Best paper relevance (2.15/3) and recent high-impact coverage (18.5%)
• Eliminates DOI hallucination entirely on this benchmark

Evidence Graph Precision

Domain-expert audit of 148 edges: 97.9% edge-type precision (143/146 decidable edges correct)

Theoretical and Practical Implications

Theoretical Implications

1. Unit of retrieval matters: Shifting from document-level to claim-level retrieval fundamentally changes how literature synthesis can be performed, enabling verifiable cross-paper answers.
2. Complementary structures over shared data: The paper demonstrates that faceted taxonomies, evidence graphs, and principle-centered hierarchies can coexist as different views over the same claim store, serving different user needs.
3. Provenance as a design principle: Every claim carries source DOI and verbatim quote, establishing traceability that helps detect unsupported generated text—an important design pattern for AI-assisted scientific workflows.

Practical Implications

1. Reduced citation fabrication: AskChem eliminates DOI hallucination on the benchmark, addressing a critical problem in LLM-based literature synthesis.
2. Agent-friendly infrastructure: REST API, SDK, and MCP server make the claim store directly usable by AI agents for grounded answer generation.
3. Interactive browsing: The faceted taxonomy supports hierarchical retrieval, grouping, and temporal exploration, not just keyword search.
4. Open infrastructure: MIT-licensed source code and CC-BY index snapshot enable community reuse and extension.

Limitations Acknowledged

• Corpus covers only a fraction of chemistry
• Abstract extraction is shallower than full-text extraction
• LLM-generated claims, relations, and taxonomy placements can be wrong
• AskChem-Bench measures groundedness on 30 questions, not full factual accuracy
• String-based taxonomy normalization can merge distinct categories or retain near-duplicates
• Retrieval gain from taxonomy-based recall has not been isolated

Conclusion

AskChem demonstrates a claim-centered alternative to document-only chemistry literature search. By making provenance-carrying claims the central unit, the system supports:

• Search results that can be verified against source quotes and DOIs
• Organization through stabilized facets
• Connection through evidence relations across papers

The deployed system operates at corpus scale (2.4M claims, 147K papers) and exposes the same claim objects to both human users and AI agents through web UI, REST API, SDK, and MCP server. On cross-paper chemistry questions, AskChem improves citation groundedness (100% resolvable DOIs vs. 88.3% without retrieval) and achieves the highest citation density (18.1 verified DOIs per answer), suggesting that claim-centered infrastructure is a practical foundation for chemistry literature synthesis.

Future directions implied by the work include: expanding corpus coverage, improving full-text extraction, refining taxonomy normalization, isolating retrieval gains from different components, and extending the approach to other scientific domains.

Availability

• Live at: https://askchem.org
• Screencast demo: https://youtu.be/SOjueOlPS-8
• MIT-licensed source code: https://github.com/bingyan4science/askchem
• CC-BY index snapshot: https://huggingface.co/datasets/bing-yan/askchem