// the find
deeplethe/utopia
World's first open-source enterprise world model.
Utopia is a self-hosted knowledge graph engine written in Rust: one binary plus Postgres (with pgvector for embeddings and a plain table as job queue) that ingests documents and structured sources into a bitemporal graph, so every fact carries both when it was true and when the system learned it. It's aimed at enterprises that want an audit-trail-grade alternative to a vector store or plain knowledge graph, with agents reading it over MCP rather than writing to it freely.
The bitemporal model is implemented properly, not just claimed — corrections close the old fact and link to the replacement instead of overwriting, so you can reconstruct what the system believed at any past point, and edges are reified so relationships can carry their own attributes. The operational footprint is genuinely minimal: full-text search is embedded (Tantivy), vectors live in pgvector, and the job queue is a Postgres table, so there's no separate infra to run beyond the binary and the database. Conflict handling is explicit rather than silent — three conflict types (data clash, axiom violation, ontology contradiction) each get distinct resolution choices, and derived facts are marked as derived and always show their premises. MCP integration exposes read-only typed queries to agents while routing writes through a human review queue, which is the right default for something billed as an audit trail.
It's v0.1 with forward-only migrations and no rollback — the README itself warns to pin versions and back up before every upgrade, which is a real constraint for anything you'd call an audit system. Nearly every non-trivial operation (extraction, entity resolution, adjudication) depends on an LLM call to a configured chat endpoint, but there's no data on latency, cost, or throughput at realistic document volumes; the 100k-document benchmark is still on the roadmap, not delivered. The architecture is single-binary/single-Postgres with no clustering or scaling story, so it's unclear how this behaves once a KB grows past a few hundred thousand facts. The README leans hard into philosophical framing (Ptolemy, Copernicus, a 'world model') for what is, underneath, a Postgres-backed RAG and entity-resolution pipeline — the actual engineering is more interesting than the marketing copy suggests, but you have to dig past it to find that out.