Ground your agents in real, cited knowledge.
Mycelium turns SPARQL queries over Wikidata into a first-class tool — so a chat or coding agent pulls real, citable facts from a public knowledge graph instead of guessing. Every answer comes back as RDF, tied to a real Wikidata entity you can trace.
A grounded answer carries its source.
An ungrounded model produces plausible-sounding facts with no provenance. Mycelium makes the knowledge graph a tool the agent can query — and cite.
Mycelium runs SPARQL against the Wikidata Query Service and returns the result as materialized RDF — Turtle by default, or N-Triples, or raw SPARQL-results JSON — with a binding count. The agent gets real entities (QIDs) and their labels, not a guess. Because the whole thing is deterministic and content-addressed, a repeated query is served from a sha256-keyed cache (from_cache) — same query, same bytes, no round-trip. It's a thin BFF over mycelium serve: it owns no data and grounding is strictly read-only.
Every fact comes back as RDF tied to a real Wikidata entity you can click. Grounding you can trace, not a plausible guess.
A named grounding over real entities.
The Groundings panel materializes a curated SPARQL query — here, programming paradigms and patterns — anchored on real Wikidata roots via wdt:P31 ("instance of"). These QIDs are real Wikidata items.
| root concept | wikidata | predicate |
|---|---|---|
| programming paradigm | wd:Q188267 | wdt:P31 |
| software design pattern | wd:Q181156 | wdt:P31 |
| architectural pattern | wd:Q635346 | wdt:P31 |
| creational pattern | wd:Q1063550 | wdt:P31 |
| anti-pattern | wd:Q76438 | wdt:P31 |
| type system | wd:Q865760 | wdt:P31 |
A catalog you extend without code.
Named groundings live in a catalog the console lists. There's one today — more are data, not a code change. Any SPARQL an agent needs it can also run directly via the tool.
| grounding | description |
|---|---|
| Programming paradigms | instances of the paradigm / pattern / type-system roots in Wikidata |
Ask → resolve → materialize → ground.
Deterministic, cached, and read-only end to end.
Ask
An agent calls mycelium__query with a SPARQL string (sparql required; format turtle | ntriples | raw_json; refresh to bypass the cache) — or the console runs a curated grounding.
Resolve
mycelium runs the query against the Wikidata Query Service — or serves it from a disk cache keyed by the sha256 of the query text. from_cache flags a hit; no round-trip.
Materialize
Results come back as RDF — Turtle by default, N-Triples, or raw SPARQL-results JSON — with a row (binding) count.
Ground
The agent gets real entities (QIDs) and labels it can cite, instead of a plausible-sounding fabrication.
Real entities beat confident fiction.
The alternative is the model's own recollection — no source, no repeatability, no way to check.
Grounding is read-only — mycelium runs your query and returns what Wikidata says. It never invents an entity.
source · github.com/fastverk/plugin-mycelium · tools: mycelium__query
Prove it's safe to merge.
mycelium is one of 14 plugins in the fastverk console — hosted, or in your own cloud.