ASTER: Agentic Text-to-SPARQL with Task Skills, Autonomous Exploration, and Refinement

We present ASTER, an open-source skill-driven agentic Text-to-SPARQL system over Wikidata. ASTER com-plements a general-purpose language-model agent with six inspectable Agent Skills: an end-to-end text2sparqlorchestrator, Wikidata entity and property search, bounded graph exploration, SPARQL generation, SPARQL execution, and optional retrieval of relevant few-shot examples. Rather than one-shot translation of a natural-language question into a SPARQL query, the agent searches for candidate identifiers, probes how the relevantfacts are represented in Wikidata, executes candidate queries, and uses endpoint results and errors to refine itsassumptions.ASTER runs on the OpenCode agent runtime and can be used interactively through the OpenCode IDEor shell, or programmatically through an OpenCode server session. During the demonstration, attendees canselect a WikiKGQA question or submit their own question, follow every skill invocation and tool call, inspectintermediate SPARQL queries and endpoint responses, and observe execution-grounded refinement in real time.Each run produces an executable SPARQL query, its answer set, a structured machine-readable session log, and aformatted Markdown trace.ASTER achieved Macro QALD F1 scores of 0.88 and 0.81 on the WikiKGQA English with-mentions andno-mentions test tracks at the Wiki-Based Knowledge Graph Question Answering (WikiKGQA) Challenge,ranking first and second, respectively. The source code with Agent Skills, execution scripts, evaluation utilities,example outputs, and a demonstration video illustrating the full flow are publicly available.

Source Code: https://github.com/IBM/agentic-text2sparql/ Demo video: https://youtu.be/vMPoaUC2vNs

Keywords: Knowledge Graph Question Answering, Text2SPARQL, Wikidata, Agent Skills, LLM Agents