Try it

Playground

Ask a question over the corporate knowledge graph, get a SPARQL query from a recorded Hermes system or from a model you run, and replay it one pattern at a time.

What runs where

Everything runs in your browser. The graph (about 26.6k triples of the Text2SPARQL corporate knowledge graph) is loaded into Oxigraph, a SPARQL engine compiled to WebAssembly. It is the engine Hermes itself uses for local graphs (through its Python binding, pyoxigraph), so the answers are the ones Hermes's own runs would get. Hermes itself is Python and is not run here: the recorded queries are what its systems wrote with Qwen3-14B, and the model prompt is a port of the training-free system's. No model is hosted for this page, so Ask my model uses a model you run: Ollama on this computer is detected automatically, or you can load a tiny model that runs in your browser on WebGPU (a few hundred megabytes, kept by your browser), or point it at another endpoint.

Loading the playground…

How to use it

  1. Pick a sample question and press one of the recorded systems: the query it wrote appears, with its recorded score. Run it on the graph and compare what the systems wrote. The prover's query is loaded first.
  2. Press Replay step by step to add the query's patterns one at a time. After each, the solution count and its change show what a planner sees as it builds a plan, and a step that empties the result points at the wrong pattern.
  3. To use a model, open Your model, point it at an OpenAI-compatible endpoint (Ollama is the default), and press Ask my model. The prompt is the training-free system's; if the draft does not parse the page asks once to repair it, as that system does.

Why Oxigraph, and not OntoDB?

Hermes's default local store is Oxigraph (see Graph stores), so the browser runs the same engine. OntoDB is not one of Hermes's stores, and it could not run Hermes's queries if it were: it answers only SELECT over a basic graph pattern with filters, and rejects OPTIONAL, UNION, property paths, and aggregates, while questions like "which bill of materials has the highest average unit cost" need aggregates. OntoDB is also a C++ engine that is not built for the web yet.

What to expect from a small model

Writing a query from this prompt is hard for small models. The tiny model that runs in your browser (LFM2.5 350M, in tests on 2026-10-22) produced a draft that did not parse, even after the one repair, for all eight sample questions, and the leaderboard shows the same trend: one-shot Qwen3-4B scores 0.000 and Qwen3-14B 0.090, against 0.282 for the best Hermes planner. Use Ollama with the largest model you have, and compare with the recorded queries above. For the in-browser model the page is mainly useful for seeing the pipeline cope: extraction, one repair, then a plain report of the failure.

What is not here