AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis

Overview§

AskChem redefines the atomic unit of chemistry literature search from paper-level retrieval to individual claims. Each paper is decomposed into a set of typed, atomic statements, each grounded by a source DOI and a verbatim quote or explicit evidence locator. This allows both humans and AI agents to directly retrieve precise findings and trace provenance without reading full PDFs.

Claim Extraction and Typing§

The claim extraction pipeline likely uses an LLM to produce atomic claims from full text. For each sentence, the model extracts:

• the claim text (subject, predicate, object, conditions)
• a type (e.g. synthesis procedure, property, mechanism, characterization)
• the evidence span (verbatim quote) and its location.

The typed claims are embedded with a dense retriever. Let $\mathbf{e}_c$ be the embedding of claim $c$ and $\mathbf{e}_q$ the embedding of query $q$. Relevance is computed as:

$$\text{score}(q,c) = \alpha \cdot \cos(\mathbf{e}_q,\mathbf{e}_c) + (1-\alpha) \cdot \text{BM25}(q,c)$$

where $\alpha$ balances semantic and lexical matching.

Complementary Structures§

AskChem organizes claims into three linked structures: 1. Stabilized faceted taxonomy – manually curated facets (reaction type, functional group, properties) for hierarchical browsing and filtering. 2. Evidence graph – entities and relations between claims, enabling path-based reasoning across papers. 3. Living taxonomy – an exploratory structure that situates papers under scientific principles and is updated as new literature is added.

These structures are all stored in a shared claim store, enabling a single query to span all access points.

Implementation and Access§

The system exposes a REST API, Python SDK, and an MCP (Model Context Protocol) endpoint for AI agents. A sample query via the SDK:

import askchem

client = askchem.AskChem()
results = client.search(
    query="yield of Suzuki coupling with palladium acetate",
    facets={"reaction_type": "cross-coupling"},
    claim_types=["synthesis_claim"],
    top_k=10
)
for r in results:
    print(r.text, r.doi, r.verbatim)

Evaluation and Impact§

On AskChem-Bench, grounding a GPT-5.5 reader with AskChem achieves 100% resolvable DOIs versus 88.3% without retrieval, and the highest citation density among five baselines. This demonstrates that claim-level retrieval reduces hallucination and improves verifiability in chemistry literature synthesis.

Takeaway§

By making claims first-class citizens, AskChem turns literature search from find documents to find evidence, significantly lowering the cost of assembling cross-paper scientific answers.