AutoSchema: Live Schema Grounding for Agentic Text-to-Sparql over Heterogeneous Knowledge Graphs

Yiming Zhang; Koji Tsuda

cs.LGAug 14, 2026

AutoSchema: Live Schema Grounding for Agentic Text-to-Sparql over Heterogeneous Knowledge Graphs

Authors: Yiming Zhang, Koji Tsuda

Organizations: The University of Tokyo Kashiwa, Chiba, Japan · Center for Basic Research on Materials, National Institute for Materials Science Tsukuba, Ibaraki, Japan · RIKEN Center for Advanced Intelligence Project Tokyo, Japan

Abstract

Life science knowledge graphs make large collections of structured data available through SPARQL, but each resource uses its own schema, identifiers, and links. TogoMCP helps language model agents query these resources by providing curated Metadata Interoperability Exchange files. Creating and maintaining these files still requires language model assisted drafting, validation, and manual review. We study live schema grounding, where an agent obtains the schema evidence needed for a question directly from the current endpoints. We present autoschema, a general framework for live schema grounding that requires no training. It inspects live schemas, maps entity names in a question to graph identifiers, explores relation paths, and finds possible connections between resources during iterative query construction. We use TogoMCP as our main comparison framework. We evaluate autoschema on Resource Focused Biomedical KGQA, Multi Resource Biomedical KGQA, Longitudinal Biomedical Semantic QA over BioASQ Task B, and Chemistry Knowledge Graph Transfer to a previously undocumented RDF graph. autoschema improves mean factoid accuracy over TogoMCP in the biomedical KGQA tasks and gives consistent gains in the longitudinal BioASQ evaluation. It also reduces iteration budget exhaustion and uses fewer tool calls on average in the core evaluation. The transfer study gives preliminary evidence that live schema grounding can support irregular and previously unseen graphs without first creating a curated schema file.

Figures & tables

Figure 1. AutoSchema in an agentic text-to-SPARQL loop over heterogeneous knowledge graphs. Given a natural-language question, the agent alternates planning, live grounding, query generation, execution, and reflection. AutoSchema supplies four question-time capabilities: source summarization, entity resolution, relation and path discovery, and cross-database bridging. Each capability returns current endpoint evidence rather than a complete schema dump. The bottom panel contrasts AutoSchema with TogoMCP and a no-schema lower bound. The left column shows a natural-language question and an LLM agent iterating through planning, AutoSchema tool calls, SPARQL generation, endpoint execution, and reflection before producing an answer. The center column details four live grounding capabilities with examples: source summarization, entity resolution, relation and path discovery, and cross-database bridging. The right column shows heterogeneous gene, disease, drug, and pathway graphs connected by possible bridges. A bottom comparison summarizes the no-schema lower bound, TogoMCP, and AutoSchema.

Task	Data and evaluation focus
Resource Focused Biomedical KGQA	RDF Portal. Factoid and list questions within one resource.
Multi Resource Biomedical KGQA	RDF Portal. Factoid and list questions across resources.
Longitudinal Biomedical Semantic QA	Six BioASQ Task B sets. Factoid, list, and yes or no questions collected over time.
Chemistry Knowledge Graph Transfer	tmQM-RDF. Factoid and list questions on a graph outside biomedicine.

Table 1. Evaluation tasks. Graph-supported means that each gold answer is verified by executing a gold SPARQL query.

Model	Framework	Factoid acc.	List F1	Limit rate
gpt-oss-120b	No schema	$0.164\pm.019$	$0.122\pm.006$	$53.3\pm 1.8\%$
	TogoMCP	$0.198\pm.009$	$0.134\pm.001$	$44.0\pm 1.0\%$
	AutoSchema	$\mathbf{0.260\pm.005}$	$\mathbf{0.199\pm.012}$	$\mathbf{33.5\pm 1.5\%}$
gemma4-31b	No schema	$0.239\pm.022$	$0.117\pm.009$	$17.6\pm 1.0\%$
	TogoMCP	$0.258\pm.008$	$0.134\pm.003$	$20.2\pm 0.8\%$
	AutoSchema	$\mathbf{0.281\pm.010}$	$\mathbf{0.145\pm.006}$	$17.1\pm 2.3\%$

Table 2. Resource Focused Biomedical KGQA with 559 questions. Results are mean

\pm

standard deviation over three runs.

Model	Framework	Factoid acc.	List F1	Limit rate
gpt-oss-120b	No schema	$0.178\pm.004$	$0.138\pm.001$	$45.7\pm 1.4\%$
	TogoMCP	$0.167\pm.024$	$0.165\pm.005$	$49.4\pm 0.3\%$
	AutoSchema	$0.203\pm.020$	$0.200\pm.019$	$\mathbf{40.0\pm 1.1\%}$
gemma4-31b	No schema	$0.110\pm.027$	$0.105\pm.010$	$25.1\pm 2.0\%$
	TogoMCP	$0.140\pm.008$	$0.125\pm.019$	$32.9\pm 1.6\%$
	AutoSchema	$\mathbf{0.216\pm.041}$	$0.134\pm.003$	$\mathbf{26.1\pm 1.6\%}$

Table 3. Multi Resource Biomedical KGQA with 226 questions. Results are mean

\pm

standard deviation over three runs.

8B (2020)
Model	Framework	Fact.	List F1	Yes/no
gpt-oss-120b	No schema	$0.468\pm.008$	$\mathbf{0.404\pm.031}$	$0.829\pm.026$
	TogoMCP	$0.510\pm.024$	$0.360\pm.027$	$0.838\pm.015$
	AutoSchema	$\mathbf{0.581\pm.040}$	$0.396\pm.022$	$\mathbf{0.860\pm.014}$
gemma4-31b	No schema	$0.614\pm.027$	$0.494\pm.013$	$0.875\pm.013$
	TogoMCP	$0.565\pm.030$	$0.466\pm.008$	$0.877\pm.004$

Table 4. Longitudinal Biomedical Semantic QA on two BioASQ Task B test sets. Results are mean

\pm

standard deviation over three runs.

Figure 2. Factoid accuracy gain of autoschema over TogoMCP. Bars show differences between three-run means. Error bars show the root sum square of the reported standard deviations and are descriptive rather than confidence intervals. Two bar charts show the change in factoid accuracy from TogoMCP to AutoSchema. The first compares two models across Resource Focused Biomedical KGQA, Multi Resource Biomedical KGQA, BioASQ Task 8B, and BioASQ Task 13B. The second shows gpt-oss-120b results across six yearly BioASQ Task B test sets. Every mean difference is positive.

Figure 3. Paired trace for a Multi Resource Biomedical KGQA question. TogoMCP continues from an incorrect static relation and reaches the interaction limit. Autoschema uses live entity resolution and local path discovery to return the correct answer. The example explains a mechanism and does not estimate how often it occurs. A question and gold answer appear above two agent traces. The TogoMCP trace plans a direct collection-number relation, issues queries that return no matches, and reaches the iteration limit. The AutoSchema trace resolves the collection record, discovers its path to a strain and phylum, and returns Actinobacteria in six calls.

Model	Framework	Latency (s)	Tool calls
gpt-oss-120b	No schema	$71.4\pm 5.6$	$13.22\pm.04$
	TogoMCP	$67.1\pm 5.9$	$12.91\pm.05$
	AutoSchema	$\mathbf{59.5\pm 11.9}$	$\mathbf{11.13\pm.12}$
gemma4-31b	No schema	$80.5\pm 1.3$	$9.31\pm.08$
	TogoMCP	$\mathbf{74.2\pm 1.6}$	$9.22\pm.05$
	AutoSchema	$76.4\pm 3.2$	$\mathbf{8.27\pm.23}$

Table 5. Mean per-question interaction cost across the two biomedical KGQA tasks and two selected BioASQ sets. Values are mean

\pm

standard deviation over three runs.

Framework	Fact.	List F1	Limit	Time (s)	Calls
TogoMCP	$0.273$	$0.167$	$35.9\%$	$\mathbf{1060.3}$	$\mathbf{15.1}$
AutoSchema	$\mathbf{0.394}$	$0.167$	$\mathbf{23.1\%}$	$1306.5$	$16.7$

Table 6. Chemistry Knowledge Graph Transfer on 39 tmQM-RDF questions with gemma4-31b. Results are from one run.

gpt-oss-120b
Task	Access	Fact.	List F1	Limit	Time (s)
R	Iterative	$0.260\pm.005$	$0.199\pm.012$	$33.5\pm 1.5\%$	$69.6\pm 17.6$
R	Single shot	$0.246\pm.013$	$0.185\pm.011$	$35.5\pm 1.0\%$	$66.3\pm 1.9$
M	Iterative	$0.203\pm.020$	$\mathbf{0.200\pm.018}$	$40.0\pm 1.1\%$	$77.2\pm 19.4$
M	Single shot	$0.185\pm.017$	$0.155\pm.010$	$41.9\pm 2.9\%$	$68.4\pm 1.6$
gemma4-31b

Table 7. Iterative and single shot grounding. Values are mean

\pm

standard deviation over three runs. R and M denote the Resource Focused and Multi Resource Biomedical KGQA tasks.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Tool and inputs	Returned grounding evidence
get_rdf_schema database , optional entity_class=""	Query scope, frequent classes with counts and example IRIs, observed properties with range kinds and values, and example SPARQL queries. The optional class name narrows the property view.
search_rdf_entity database , keyword , optional limit=10	Candidate entity IRIs and labels found through common label and identifier properties. For recognized biomedical accessions, the response may provide a directly constructed IRI instead of a label match.
expand_entity database , entity_iri	Direct properties and paths that pass through as many as three nested blank nodes. Each nested path includes a SPARQL pattern that can be run again to obtain a filterable or joinable result.
get_cross_db_bridges database_a , database_b	Candidate links based on an optional identifier conversion service, direct IRI reuse, or property names observed in both sources. The response includes a join skeleton when a shared property is found.

Appendix

Table 8. Core MCP tools exposed by autoschema . Defaults are shown in the signatures.

Natural Language to SPARQL

Natural language to SPARQL translates questions into executable queries over knowledge graphs.

See all 15 papers

Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval
Jun 3, 2026Padmaja Jonnalagedda, Yuguang Yao, Xiang Gao +2KG Construction Knowledge Representation
Intuit AI Research
SAGA: Schema-Aware Grounding for Agentic Text-to-SPARQL Generation
Jul 16, 2026Yiming Zhang, Koji TsudaLLM Grounding Knowledge Graph Question Answering
1The University of Tokyo, Japan · Center for Basic Research on Materials, National Institute for Materials Science, Japan · 3RIKEN Center for Advanced Intelligence Project, Japan
Enabling Knowledge Graph Understanding at Scale with the EXplore Your Graphs ENgine (EXYGEN)
Sep 10, 2026Harshdeep Singh, Yurui Zhu, Giovanni Colavizza +1Graph Retrieval-Augmented Generation Knowledge Graph Question Answering
Odoma Ltd.

Jun 3, 2026J/K move · Enter open · S save

Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval

Padmaja Jonnalagedda, Yuguang Yao, Xiang Gao, Hilaf Hasson+1

Intuit AI Research

KG Construction Knowledge Representation Knowledge Graph Question Answering Hybrid Retrieval Data Integration

Open paper AskPDF arXiv

Real-world data spans tables, documents, and semi-structured files with implicit semantics. Querying this data requires integrating evidence across inconsistent schemas and formats, yet existing approaches either demand costly manual engineering or bypass structure entirely. We present a system that automatically discovers an executable schema from raw multi-source data and uses it as a shared contract for knowledge graph construction and query-time retrieval. A closed-world field catalog constrains LLM-based schema discovery to attested fields; deterministic structural analysis infers identity keys, foreign keys, and source hierarchy; and the resulting schema drives extraction, deduplication, and cross-source linking into a provenance-aware knowledge graph. At query time the schema -- optionally extended via a monotonic protocol -- conditions a multi-tool agent routing retrieval across structured lookup, graph traversal, and vector search, returning grounded answers with traceable citations. In controlled zero-shot comparisons using the same LLM, data, and evaluation harness, the system improves over retrieval-only and decomposition-based baselines across four QA benchmarks, with ablations showing that schema-conditioned routing, structural intelligence, and schema-guided construction each contribute to the gains.

Abstract

Natural Language to SPARQL

Explore similar work

Executable Schema Contracts: From Automatic Ingestion to Multi-Source Retrieval

SAGA: Schema-Aware Grounding for Agentic Text-to-SPARQL Generation

Enabling Knowledge Graph Understanding at Scale with the EXplore Your Graphs ENgine (EXYGEN)