AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Analysis

Why This Paper Matters

AskChem addresses a critical bottleneck in scientific literature synthesis: the mismatch between document-level retrieval and the need for precise, cross-paper evidence assembly. Traditional search systems return ranked lists of papers, forcing scientists and AI agents to manually locate relevant passages, verify provenance, and synthesize answers. By shifting the retrieval unit to atomic, provenance-carrying claims, AskChem enables a more direct and reliable path from query to evidence.

This work is particularly timely given the rise of large language models (LLMs) and AI agents that rely on retrieval-augmented generation (RAG). The authors demonstrate that grounding an LLM in a claim-level store significantly improves citation resolvability, a key metric for trustworthiness. The infrastructure's design—combining a faceted taxonomy, evidence graph, and living taxonomy—offers a blueprint for building domain-specific knowledge infrastructures that support both human and machine users.

Technical Contributions

• Claim-centered retrieval: Each paper is decomposed into atomic, typed claims, each with a source DOI and a verbatim quote or evidence locator, ensuring provenance and verifiability.
• Shared claim store: Indexes 2.4M claims from 147K papers, providing a unified resource for search and synthesis.
• Complementary structures: A stabilized faceted taxonomy for hierarchical retrieval, an evidence graph linking claims through relations, and a living taxonomy that situates papers under scientific principles.
• Multi-modal access: Web interface, REST API, SDK, and MCP (Model Context Protocol) support for AI agents, making the infrastructure easily integrable into existing workflows.
• AskChem-Bench: A benchmark for evaluating claim-level retrieval and grounding, enabling standardized comparison of systems.

Results

On AskChem-Bench, grounding a GPT-5.5 reader in AskChem achieves 100% resolvable DOIs, compared to 88.3% without retrieval. This indicates that the claim-level retrieval provides accurate and complete source attribution. Additionally, AskChem achieves the highest citation density among five tested systems, suggesting that its responses are more densely grounded in verifiable sources. These results highlight the practical benefits of claim-centered retrieval for improving the reliability of AI-generated scientific answers.

Significance

AskChem represents a significant step toward making scientific literature synthesis more efficient and trustworthy. By providing a claim-level infrastructure with robust provenance, it enables both human researchers and AI agents to assemble cross-paper answers with confidence. The design principles—atomic claims, explicit provenance, and multiple organizational structures—could be adapted to other scientific domains, potentially leading to a new generation of literature search systems that prioritize evidence quality over document ranking. As AI agents become more prevalent in research, infrastructures like AskChem will be essential for ensuring that their outputs are grounded in verifiable facts.