AskChem: Making Provenance-Carrying Claims the Retrieval Unit

The paper in 90 seconds

• Problem: Conventional scientific literature retrieval treats entire documents or arbitrary-length text chunks as its fundamental unit. When researchers or autonomous agents attempt to synthesize findings across dozens of papers, they must manually locate supporting sentences within verbose passages, ascertain whether an assertion is genuinely traceable, and assemble cross-paper relationships by hand. This paradigm frequently causes citation disconnection, contextual truncation, and post-hoc citation hallucination.
• Core insight: AskChem elevates the basic retrieval and knowledge unit from raw documents to atomic, typed scientific claims. Each claim carries a durable identity, is permanently bound to a source DOI and a verbatim quote or evidence_locator, and is organized through an induced faceted taxonomy across multiple operational views alongside an evidence graph capturing cross-document relations.
• Strongest evidence: In a corpus of 2.4 million claims, 147,000 papers, and 171,342 typed evidence edges, AskChem-Bench evaluated 30 cross-paper chemistry synthesis questions. GPT-5.5 grounded with AskChem achieved a 100% DOI resolvability rate in CrossRef (versus 88.3% for LLM-only) and averaged 18.1 verified citations per answer (versus 9.6 for LLM-only) (Section 7; Table 1).
• Main boundary: DOI resolvability is a metric of citation plumbing and traceability; it does not prove that extracted claims are semantically correct or that cited statements constitute scientific truth. Furthermore, on grounded quantitative detail (grounded specificity), the specialized research agent Edison Scientific markedly outperformed AskChem (29.2 versus 5.9), and the paper provides no isolated component ablations.

When an answer to a research question is scattered across dozens of papers, receiving a list of paper titles or raw text chunks leaves the hardest work—localization, verification, and cross-literature reconciliation—entirely to the reader or the downstream agent. AskChem’s core research question is: can a scientific claim, carrying its immutable provenance and structured conditions, become the first-class infrastructure object searched, browsed, linked, and reused across platforms?

This reading is based on the arXiv cs.CL v1 preprint by Bing Yan, Gregory Wolfe, Stefano Martiniani, and Kyunghyun Cho, submitted on 2026-07-30 (arXiv:2607.28618v1). Figures 2 through 5 are reproduced from the open paper.

What to know first

To evaluate AskChem’s system architecture, one must first recognize the structural bottlenecks of prior approaches when navigating scientific literature, as well as the foundational data models introduced by the paper:

Bottlenecks of prior approaches: why traditional paper-ranking and chunk-level RAG fall short

1. Document-level granularity mismatch: Traditional scientific search engines return full PDF papers or multi-page abstracts. The synthesis model must search through thousands of words to determine which specific sentence supports which claim. As a result, citations frequently remain at the paper-title level, making rapid sentence-level auditing difficult.
2. Chunk-level RAG fragmentation and phantom citations: Splitting text into arbitrary chunks (e.g., 512 or 1024 tokens) frequently severs necessary experimental conditions from their conclusions. For instance, if an essential catalyst preparation step or an additive is isolated in one chunk while the resulting 98% efficiency appears in another, retrieval may miss the critical constraint. Furthermore, when generative models synthesize an answer, they often generate citation markers as statistical tokens rather than traceable data pointers, leading to invalid or hallucinated citations.

AskChem’s core data abstractions

• Atomic typed claim: The minimal self-contained scientific assertion extracted from literature. Each claim is assigned a unique claim_id and structured fields (reaction type, substance, yield, quantitative values, units, and preconditions), an extraction model identifier and confidence score, and an inseparable provenance anchor.
• Provenance locator: Every claim is strictly bound to its source DOI. When the evidence consists of a contiguous passage, AskChem attaches a verbatim quote; when the finding spans discontinuous paragraphs, tables, or supplementary artifacts, it uses a structured evidence_locator.
• Faceted taxonomy: Rather than forcing all chemistry knowledge into a single rigid ontology tree, the system induces nine orthogonal operational views (Reaction, Substance, Application, Technique, Mechanism, Data, Claim Type, Time, and Author), enabling the same claim to be discovered from multiple perspectives.
• Typed evidence graph: A cross-document network capturing structured relationships between claims, including supports, contradicts, extends, derives_from, and cites_as_evidence.

Core intuition

AskChem’s central contribution is a fundamental shift in the decision rule and mental model of literature retrieval systems:

Previous decision rule (Document / Chunk retrieval):
Retrieve coarse documents or chunks → LLM reads context and freely generates conclusions → LLM attempts to append citations post-hoc
(Citations are post-hoc generated tokens prone to hallucination; difficult to verify whether the source actually supports the assertion)

AskChem decision rule (Claim-centered retrieval):
Offline pipeline extracts atomic claims with immutable provenance → Claims are indexed into multi-view taxonomy and evidence graph → Retrieval outputs candidate evidence bundles directly
→ Synthesis model compares and aggregates across pre-anchored claim objects
(Candidate evidence is traceable by construction; provenance validation is shifted upstream into the data model)

This transforms the paradigm from “asking the generator to prove its citations after the fact” to “ensuring all candidate evidence carries verified provenance anchors before entering the synthesis prompt.”

However, an essential defensive intuition must be maintained: changing the retrieval unit resolves the citation plumbing problem, but it does not eliminate extraction errors. If an offline extractor misinterprets a source sentence, overlooks a negative qualifier, or misidentifies a unit of measurement, that claim will still enter the index with a valid DOI. AskChem provides an auditable lever that allows readers or verification agents to jump back to the exact source text with one click; it does not constitute an automated, unsupervised arbiter of scientific truth.

Walk one example through the method

To trace the complete pipeline in practice, consider a canonical electrochemical carbon dioxide reduction (CO₂ reduction) inquiry from Appendix B:

1. Input (synthesis question): A researcher or autonomous scientific agent asks: “Under what conditions and temperature range does single-atom Ni catalysts reduce CO₂, what Faradaic efficiency and turnover frequency (TOF) are reported, and do different papers report contradictory findings?”
2. Intermediate representation (subquery expansion and hybrid recall):
   • The system expands the query into three to four keyword subqueries and executes a parallel fan-out across hybrid retrieval channels.
   • The retriever returns multiple matching atomic claim objects. For example, it retrieves claim_id: 7c92fcacd8cb64d4, linked to DOI 10.1002/anie.201914977, recording that catalyst Ni SA-N2-C in CO₂ reduction achieved 98% CO Faradaic efficiency and a TOF of 1622 h⁻¹, complete with its verbatim source quote.
3. Decision or transformation (taxonomy clustering and graph navigation):
   • Via the Faceted Taxonomy, the claim is mapped to operational nodes such as reaction/co2_reduction and substance/single_atom_catalyst.
   • Traversing the Evidence Graph, the system identifies a related paper (DOI: 10.1039/D0EE01234A) claiming that under comparable overpotentials, the primary reduction product was formic acid (HCOOH) rather than CO, which the graph flags as a potential condition divergence or contradiction edge.
4. Output (grounded synthesis generation):
   • The synthesis reader (GPT-5.5) aggregates findings across 40 diversified candidate claims. In the generated output, every quantitative performance metric and reaction condition is tied to its explicit DOI and source quotation, allowing the user to click any claim to inspect the source paper.
5. Likely failure point (potential failure modes):
   • If the extraction model missed an exclusion clause in the original text (such as “Ni SA-N2-C did not exhibit high activity in the absence of nitrogen coordination”), the resulting claim will present an inverted finding. The DOI will still resolve cleanly, but unless an auditor verifies the quote, the semantic error will persist into the summary. Additionally, if taxonomy fuzzy clustering force-fits an outlier mechanism into an improper category node, downstream relational traversal will misfire.

Technical mechanism

AskChem’s end-to-end architecture is built on an industrial-scale data processing and retrieval pipeline comprising five major stages:

1. Dual-track extraction pipeline: A high-throughput pipeline utilizes GPT-5-mini to process titles and abstracts, while a deeper full-text pipeline uses Gemini 3.1 Pro with native PDF inputs and Vertex AI batching to extract complex assertions from document bodies, figure legends, and appendices.
2. Schema validation and retry gates: Every extraction output is validated against strict Pydantic schemas enforcing required provenance fields (DOI, verbatim quotes, locator coordinates), numeric validity ranges, and chemical naming standards. Schema-invalid outputs trigger automated prompt retries.
3. Multi-view relational data model: Using a unique claim_id as the primary key, SQLite/relational databases maintain synchronized alignments across a Source table (DOI, publication year, citation counts, OpenAlex author metadata), a TreeNode table (multi-view classification paths), and an Edge table (typed claim-to-claim relationships).
4. Four-way hybrid retrieval with RRF: The search engine fuses SQLite FTS5 full-text search, paper-level BM25 recall, taxonomy-node recall, and dense vector embeddings using Reciprocal Rank Fusion (RRF) to merge candidate rankings.
5. Cross-paper relation extraction and interfaces: For co-occurring and topically aligned claims, a secondary relational extraction pass identifies cross-paper dependencies; the resulting knowledge substrate is exposed through a web interface, REST API, Python SDK, and an MCP server.

Figure 2 shows the claim as retrieval unit connected to a faceted taxonomy, evidence graph, and Living Taxonomy. Paper Section 1. Source: AskChem Figure 2, Bing Yan et al.; used under the paper’s CC BY 4.0 license, checked 2026-08-07.

Three complementary organizing structures

Rather than attempting to construct an all-encompassing monolithic ontology, AskChem divides the organization of scientific claims into three complementary structures:

1. Faceted taxonomy

Section 4 explains that category paths are dynamically induced during paper and claim ingestion, then stabilized through canonical top-level routing, synonym normalization, and fuzzy clustering into persistent L1/L2/L3 hierarchical paths.

Figure 4 shows CO₂-reduction claims through reaction, substance, application, technique, mechanism, data, claim type, time, author, and network views. Source: AskChem Figure 4, Bing Yan et al.; used under CC BY 4.0, checked 2026-08-07.

The faceted taxonomy allows a single CO₂ reduction claim to be filtered simultaneously by reaction path, single-atom catalyst substance, or in-situ spectroscopy technique. These paths serve directly as recall channels; however, because the authors did not conduct an independent expert audit of taxonomy placement accuracy, paths should be understood as operational grouping heuristics rather than validated scientific ontologies.

2. Cross-paper evidence graph

Section 3 reports 171,342 typed relation edges across the corpus. The authors conducted a domain-expert audit on a stratified sample of 148 edges. Excluding two undecidable cases, 143 of 146 relation types were judged correct, yielding a 97.9% edge-type precision.

Figure 3 summarizes corpus coverage and automatic quality checks for the deployed index. Its caption also warns that these statistics do not replace expert judgments of claim semantics or taxonomy placement. Source: AskChem Figure 3, Bing Yan et al.; used under CC BY 4.0, checked 2026-08-07.

It is vital to distinguish that 97.9% edge precision measures whether the relation label correctly describes the relationship between two identified claims; it does not measure the graph’s overall recall across the literature, nor does it guarantee the factual accuracy of the connected claims themselves.

3. Principle-centered Living Taxonomy

While the faceted taxonomy serves everyday search and filtering, the Living Taxonomy attempts an exploratory organization of literature based on underlying scientific principles, theories, models, mechanisms, and phenomena. Section 5 and Appendix B Table 3 document 4,931 nodes, approximately 1.1 million claims, and 360,546 paper placements, including 663 open proposed branches.

Figure 5 is a screenshot of the principle-centered Living Taxonomy. Paper Section 3. Source: AskChem Figure 5, Bing Yan et al.; used under CC BY 4.0, checked 2026-08-07.

The authors explicitly characterize the Living Taxonomy as exploratory. Because nearest-neighbor vector embeddings guide placement, low-margin boundary cases can be force-fitted into unsuitable host concepts. It is valuable for serendipitous literature exploration, but should not be relied upon as an unsupervised decision arbiter.

How to read the evidence

Evaluating a scientific literature synthesis system requires scrutinizing the benchmark tasks, baseline configurations, evaluation metrics, and experimental controls across all key dimensions.

Experimental setup and controls

• Benchmark dataset and tasks (datasets): AskChem-Bench v1.1 comprises 30 curated cross-paper synthesis questions divided equally among three distinct tasks (10 questions each):
  1. Cross-Paper Condition Aggregation (CA): Aggregating catalysts, operational conditions, and performance metrics across disparate studies to evaluate quantitative detail assembly.
  2. Temporal Claim Tracking (TC): Tracing the evolution of scientific concepts over time to test whether chronological progression is preserved.
  3. Contradiction Surfacing (CS): Retrieving and exposing conflicting experimental findings to evaluate divergence identification.
• Comparison baselines (baselines): All evaluated settings utilized GPT-5.5 as the synthesis reader under five distinct retrieval configurations:
  1. LLM only: Unaugmented model relying solely on internal parameters.
  2. +AskChem: Grounded with AskChem’s claim-level hybrid retrieval bundle.
  3. +Paperclip: Grounded with conventional paper-level retrieval.
  4. Edison Scientific: A specialized commercial scientific research agent (PaperQA architecture).
  5. NotebookLM Deep Research: Google NotebookLM’s deep literature synthesis workflow.
• Evaluation metrics (metrics):
  • DOI existence (%): The percentage of cited DOIs that resolve successfully in CrossRef.
  • Citation density (/answer): The average count of distinct, verified DOIs cited per response.
  • Grounded specificity: The count of quantitative tokens appearing in the same sentence as an explicit citation marker.
  • Recent high-impact (%): The proportion of cited papers published within the past 5 years with 50 or more citations.
  • Paper relevance (0–3): Relevance scored by Gemini 3.1 Pro (3: direct, 2: on-topic, 1: loose, 0: irrelevant), calibrated against 100 domain-expert annotations with 93% agreement ($\kappa = 0.914$).
  • On-topic ≥ 2 (%): The proportion of cited papers receiving a score of 2 or 3.
• Compute and retrieval controls (compute): AskChem utilized query rewriting to fan out across 3–4 subqueries, capping diversified candidate evidence at 40 claims. For full-text extraction, Vertex AI batch processing with Gemini 3.1 Pro was deployed. Because Edison and NotebookLM operate as closed systems with differing tool budgets and retrieval depths, the comparison reflects an end-to-end profile comparison rather than an isolated compute-controlled shootout.

Table 1: experimental results reproduction

The table below reproduces the complete evaluation results from Table 1 of the paper (overall mean across 30 benchmark questions):

Metric	LLM only	+AskChem	+Paperclip	Edison Scientific	NotebookLM
DOI existence (%)	88.3	100	100	99.1	93.7
Citation density (/answer)	9.6	18.1	7.5	10.7	7.9
Grounded specificity	8.1	5.9	0.5	29.2	0.1
Recent high-impact (%)	0.6	18.5	6.1	11.3	12.1
Paper relevance (0–3)	1.66	2.15	1.72	2.07	1.84
On-topic ≥ 2 (%)	65.8	86.6	57.8	89.7	78.9

Interpreting the evidence and its boundaries

1. Dramatic reduction in citation hallucination: §7 RQ3 and Table 1 show that grounding GPT-5.5 with AskChem raised DOI existence from 88.3% to 100% and nearly doubled citation density from 9.6 to 18.1. In the representative case ca04 (Figure 6 / §7), the LLM-only model generated 14 DOIs of which 6 were unresolvable fabrications, whereas the AskChem-grounded synthesis produced 22 citations that all resolved cleanly. This directly supports the claim that claim-centered grounding fixes citation plumbing in literature synthesis.
2. Trade-offs in quantitative grounded specificity: Table 1 reveals that for grounded specificity (quantitative tokens per cited sentence), Edison Scientific scored 29.2, far surpassing +AskChem’s 5.9. The authors acknowledge that Edison extracts substantially denser numerical parameters from full texts. AskChem’s architecture excels at citation traceability, but it does not lead across all dimensions of informational detail.
3. Relevance scores and coverage limits: +AskChem achieved the highest mean relevance score (2.15 versus Edison’s 2.07), yet Edison achieved a slightly higher on-topic rate of 89.7% versus AskChem’s 86.6%. The results demonstrate comparable top-tier retrieval quality, rather than an unmitigated victory across all settings.
4. Lack of isolated component ablations: The evaluation lacks an ablation study isolating individual components (e.g., measuring performance when removing taxonomy recall, the vector channel, or RRF). Consequently, one cannot determine how much of the performance gain is attributable to claim indexing versus subquery fan-out and candidate diversification.

Evidence map

To maintain clear scientific hygiene, the paper’s findings and claims are divided into four distinct epistemic tiers:

Tier	What can be safely established	What must not be claimed
Direct paper evidence	- The corpus contains 2.4M claims, 147K papers, and 171,342 typed edges. - Stratified audit of 148 edges demonstrated 97.9% relation-type precision (143/146 decidable). - On 30 AskChem-Bench questions, +AskChem achieved 100% DOI existence and 18.1 verified citations/answer. - Gemini relevance evaluator showed 93% agreement ($\kappa = 0.914$) with expert calibration. - Edison Scientific demonstrated higher grounded specificity (29.2 vs 5.9) and higher on-topic proportion (89.7% vs 86.6%).	- Does not establish that all 2.4M extracted claims are semantically accurate. - Edge precision does not establish complete relational recall across literature. - Does not establish performance on large-scale tasks beyond the 30 curated questions. - High relevance scores do not guarantee experimental safety in a laboratory. - Does not establish superiority over specialized agents across every metric.
Author causal claims	- Indexing atomic claims resolves citation hallucinations at the root. - Faceted taxonomy and Living Taxonomy effectively empower researchers and autonomous agents. - AskChem serves as the premier knowledge substrate for chemistry synthesis.	- Component contributions are unisolated due to the absence of ablations. - Living Taxonomy exhibits force-fitting of low-margin concepts into host nodes. - Cross-domain generalization outside chemistry remains unproven.
Unsupported claims	- 100% DOI existence $\neq$ 100% scientific truth (traceability is not factual validity). - The architecture has not been proven effective on non-chemistry domains. - Production feasibility (update latency, indexing cost, maintenance overhead) remains unestablished.	- Do not claim that claim-centered retrieval is a universal solution for all sciences. - Do not claim that human verification can be eliminated. - Do not equate an academic prototype with an enterprise production service.
Bloss0m engineering synthesis	- Moving the retrieval unit to persistent, provenance-carrying atomic claims is a superior paradigm for verifiable RAG. - Citation plumbing (traceability) and semantic correctness must be engineered as separate verification gates. - Production systems should decouple relation validation from claim auditing and implement abstention mechanisms.	- Avoid copying large chemistry ontologies; adapt the data abstractions and validation gates. - Treat the infrastructure as an auditable evidence layer, not an autonomous oracle.

Artifacts and reproducibility

As of 2026-08-07 (with monitoring through 2026-08-09), the availability and reproducibility of AskChem artifacts are summarized below:

Artifact	Direct endpoint status	Reproducibility reading and boundaries
Live system	askchem.org is publicly accessible, offering web search and API documentation.	Usable for exploration and queries; production SLAs, caching tiers, and concurrency limits are not published; do not infer enterprise availability.
MIT source	GitHub repository is public, containing src/, sdk/, mcp_server.py, tests, and a Dockerfile under an MIT license.	Usable for inspection and code reuse; the README explicitly notes that self-hosting does not replicate askchem.org’s proprietary operational configuration, and there is no pinned release tag.
Index snapshot	Hugging Face dataset is available, containing claims.jsonl, sources.jsonl, and an approximately 25.44 GB SQLite askchem.db (totaling ~40.9 GB).	Partially usable; downloading the dataset requires significant bandwidth, and dataset previewers have noted schema-casting warnings; this reading does not claim a completed local rebuild of the full database.
Benchmark dataset	Public benchmark endpoint returns JSON directly, containing version 1.1 with 30 questions, task tags, and evaluation protocols.	Usable benchmark artifact; reproducing results requires pinning prompts, model checkpoints, and external CrossRef query timing.
REST / OpenAPI	API documentation endpoint responds with 200; search, claim, and graph endpoints return structured JSON.	Usable for bounded reproduction; users must account for anonymous rate limits and version divergence between /api/ and /v1/.
SDK and MCP	GitHub sdk/ directory, PyPI askchem package, and the MCP client are accessible.	Published and inspectable; code is ready for structural integration, but long-term maintenance and production stability remain unmeasured.

Direct queries to the /api/stats endpoint during the study period reported 2,442,810 claims, 146,627 sources, and 10,327 nodes, showing minor discrepancies from the paper’s rounded 2.4M claims, 147K papers, and 307K taxonomy nodes. In the absence of an official manifest, these snapshot numbers should not be conflated.

Bounded reproduction path: The most reliable independent verification path is to consume the 30 benchmark questions via the public JSON endpoint, target the live AskChem search API, evaluate GPT-5.5 outputs under LLM-only and AskChem-grounded prompts, and verify generated DOIs against the CrossRef API. Replicating the full-scale extraction and continuous indexing pipeline would require unreleased batch infrastructure and substantial compute resources.

Bloss0m engineering judgment and when not to use it

Drawing from the evidence and operational boundaries, Bloss0m provides the following architectural guidance:

When to adopt this architecture

1. High-stakes scientific and regulatory retrieval: In medicinal chemistry, materials science, patent analysis, or legal compliance, every generated claim must cite an exact, auditable source sentence that human auditors can verify immediately.
2. Multi-document RAG suffering from severe citation hallucination: When conventional chunk-based RAG generates fluent but untraceable citations, shifting the retrieval unit to atomic claims with persistent provenance locators directly resolves citation plumbing defects.
3. Autonomous scientific agents requiring structured inputs: When agents perform hypothesis testing, condition aggregation, or contradiction surfacing, structured claim objects with explicit condition and value attributes are far easier to manipulate logically than unstructured text chunks.

When not to use it (failure modes and boundary conditions)

1. Automated “truth oracles”: Never treat AskChem’s claim locator as a guarantee of scientific truth. A resolvable DOI merely proves that the author wrote the sentence, not that the experimental conclusion is replicable or that it has not been debunked by subsequent research.
2. Latency- and cost-sensitive consumer applications: Offline claim extraction, schema validation, graph construction, and multi-channel hybrid search incur substantial indexing costs and noticeable query latency. For general FAQ or customer service bots, this constitutes unnecessary over-engineering.
3. Niche domains lacking specialized extractors: Without domain-adapted extraction prompts and schema validation gates, general LLMs frequently misread specialized units, omit reaction conditions, or invert negation clauses, polluting the downstream index.
4. Volatile corpora without versioned governance: If documents update frequently without incremental graph reconciliation and contradiction handling, continuously ingesting uncurated claims will degrade the taxonomy and corrupt relation edges.

Production engineering blueprint

Teams seeking to implement claim-centered retrieval in production RAG systems should follow four core principles:

• Principle 1: First-class claim abstraction. Define a durable data schema containing claim_id, source_doi, verbatim_quote, conditions, confidence, and extractor_version.
• Principle 2: Decouple provenance from semantic truth. Implement two independent verification gates: an upstream locator gate (verifying that the citation and quote exist in the original text) and a downstream semantic gate (evaluating whether the claim is scientifically sound).
• Principle 3: Operational facets over rigid ontologies. Prioritize operational views tailored to specific user queries (e.g., filtering by experimental technique, date, or data type) rather than attempting to build an exhaustive, fragile global ontology.
• Principle 4: Explicit abstention on ambiguity. When an extractor or taxonomy classifier encounters low confidence, mark the entry as “requiring review” rather than allowing the model to hallucinate a category or relation.

This framework complements Bloss0m’s broader research path: RAG vs GraphRAG examines the macro-level trade-offs between chunks, global graphs, and hybrid retrieval; AskChem moves one layer deeper, addressing what fundamental data unit should be retrieved. For agents routing retrieved data into external toolkits, compare these findings with RAG-MCP to prevent schema bloat in context windows.

Three things to remember

1. Technical idea: AskChem replaces documents and raw chunks with atomic, typed scientific claims bound to source DOIs and verbatim quotes, organizing them across multi-view faceted taxonomies and cross-paper evidence graphs.
2. Evidence: On the 30-question AskChem-Bench, claim-centered grounding achieved 100% DOI resolvability and a citation density of 18.1; however, specialized systems like Edison Scientific extracted substantially denser quantitative parameters (29.2 vs 5.9), demonstrating that citation traceability does not guarantee superiority across all quality axes.
3. Boundary: Provenance is not scientific truth. Engineering implementations must maintain strict separation between citation auditability, index updating, and semantic verification, never treating a claim locator as an automated label of factual correctness.

Primary sources

• AskChem arXiv record: Title, authors, preprint version, and official abstract.
• AskChem full paper in arXiv HTML: Complete paper text covering Section 2 (claim representation), Sections 3–5 (taxonomy and evidence graph), Section 7 (evaluation), Figures 2–5, and Appendices A–B.
• AskChem PDF: Complete 10-page preprint including evaluation tables and appendices.
• AskChem source repository: MIT-licensed source code for the retrieval pipeline, Python SDK, MCP server, and Docker deployment configs.
• AskChem index snapshot: Public Hugging Face dataset containing claims, sources, and the 25.44 GB SQLite database.
• AskChem-Bench JSON and OpenAPI docs: Public benchmark question set and REST API specifications.
• CC BY 4.0: The open license governing scholarly reproduction and attribution of arXiv Figures 2–5.