Introducing Quagga Agent: Asking SSH Knowledge Graphs Questions in Plain Language
Authors: Matteo Romanello and Harshdeep Singh
Last autumn, we invited the GRAPHIA community to help us build Quagga – a crowdsourced benchmark of natural language questions and their corresponding SPARQL queries, meant to capture how scholars in SSH actually ask questions of knowledge graphs. Thanks to dozens of contributions from the community, that benchmark has grown steadily ever since. From September 2025 to June 2026, we collected over 250 question-query pairs, spanning 28 different SSH knowledge graphs and contributed by 47 different users.
Today we’re closing the loop: introducing Quagga Agent, a conversational AI agent that puts the benchmark’s purpose into practice. Instead of learning SPARQL, researchers can now simply ask a question in plain language and get a verified answer back, together with the exact query that produced it.
From Benchmark to Agent
Quagga Agent is built on top of the same knowledge graphs the community has been mapping through the Quagga platform. It takes a natural-language research question, figures out which knowledge graph is relevant, explores that graph’s schema on the fly, generates and runs a SPARQL query, and returns a structured, verified answer. Every step is visible: the agent never hides the query or the reasoning behind it.
Under the hood, Quagga Agent is a multi-agent system built with LangGraph. A main routing agent identifies the right knowledge graph for a given question and hands it off to a specialised SPARQL subagent, which works in two phases: first exploring the graph’s structure (entity types, properties, relationship patterns), then iterating on query generation until it has a validated answer. A dedicated validation step guarantees that the final SPARQL query, together with its results, is always surfaced to the user—so every answer remains scientifically transparent and reproducible.
Seeing It in Action
We recently put Quagga Agent through its paces on musicological data as part of the HERMES Data Challenge, querying the Culture Knowledge Graph (CKG) Wikidata, the German National Library (DNB), and the Digital Image Archive of Medieval Music (DIAMM). One example, developed entirely through conversation, shows what this looks like in practice: building a cross-graph profile of Johann Sebastian Bach:
- What sources by Bach exist in the Culture Knowledge Graph, and what incipits do they contain?
- In Wikidata, what are Bach’s life dates, image, occupations, and family relationships?
- What musical entries exist for Bach in DNB, and what are their instrumentation and key designations?
- Build a comprehensive profile of Bach spanning sources, biography, and published editions.
Four questions, four different knowledge graphs, zero lines of SPARQL typed by hand—yet every query the agent ran along the way remains fully inspectable. You can browse a static export of the full conversation here.
The same architecture that produced this musicological deep-dive is what now powers Quagga Agent for the broader GRAPHIA community—the underlying agent was originally developed within GRAPHIA as a general-purpose SPARQL agent for SSH knowledge graphs, and the HERMES challenge became an opportunity to stress-test it on a new domain. Readers curious about the internals—architecture diagrams, the exploration pipeline, and known limitations—can find the full technical note here.
Try it Yourself
Quagga Agent is now live at quagga.graphia-ssh.eu/chat. You’ll need to be logged in with your existing Quagga-bench credentials (GitHub, ORCID, or OPERAS ID) to access it—the same account you may already use to contribute question-query pairs to the benchmark.
Quagga, in Motion
Beyond the HERMES challenge, Quagga has been kept busy on a few other fronts. It has been put to work in a couple of “sprints” tied to Pelagios Registry activities, exploring what the agent can surface from linked open data published by Pelagios partners and members. And this September, a hybrid poster on the Quagga ecosystem—benchmark and agent together—will be presented at AI4LAM’s Fantastic Futures 2026: Trust in the Loop, taking place September 16–17 in Washington, D.C. and online.
The Loop Continues
Quagga Agent doesn’t replace the benchmark effort—it depends on it, and it will keep depending on it. Every knowledge graph you help us document, every question-query pair you contribute, makes the agent more capable across a wider range of SSH resources. If you haven’t yet, explore Quagga and consider adding your own knowledge graph or contributing question pairs in your area of expertise. The more the community puts in, the more the agent can give back.