Ask GPTKB v2

Ask a question in plain English — an LLM agent explores GPTKB v2, writes a SPARQL query that answers it, and shows you both the result and the query.

Try:

Agent trace

The agent is GRASP (Walter & Bast, ISWC 2025), self-hosted against GPTKB v2's RDF export. Your question goes to it (openai/gpt-oss-120b) — its only knowledge of GPTKB v2 comes through tool calls, and every step it takes streams here, live. Its loop:

  1. Search

    It finds candidate entities and properties in search indices over GPTKB v2's canonical labels and observed surface forms — the same disambiguation evidence the browser shows on entity pages.

  2. Verify

    It checks every E/P/C identifier with exploratory SPARQL before using it — GPTKB v2's ids are unknown to language models, so nothing is answered from model memory.

  3. Execute

    When a query produces the answer, it runs one final time — the result table and SPARQL rendered below are the ground truth, and you can rerun or edit the query in the editor.