Ask your average enterprise RAG system a simple question — "what did we think of this vendor last year?" — and it falls flat. Vector stores hold present knowledge. They have no sense of time, no record of where a fact came from, and no trace of how a conclusion changed.
Utopia (2.9k stars, Apache-2.0, Rust) takes a different route. It calls itself the first open-source enterprise world model — a knowledge system that learns passively and governs itself.
The core idea: bitemporal knowledge graph
The heart of Utopia is a bitemporal knowledge graph. Every fact carries two timelines:
- World time — when something was true in the real world;
- System time — when your system came to believe it.
When a vendor's credit rating drops from AA to A, a normal store overwrites the old record. Utopia closes the old fact, links the new one to it, and keeps both timelines intact. You can rewind to any point and see exactly what you knew then, and why.
For an enterprise, that's an auditable decision trail — not a nice-to-have, but the difference between "we decided X" and "we can prove what X was based on."
A product, not a library
Utopia ships as one Rust binary plus one Postgres. Install it and it works. The built-in feature set is dense:
- Ingest — PDF, DOCX, PPTX, XLSX, Markdown, HTML, plus scheduled sync from web, RSS, GitHub, Jira, and S3;
- Search & chat — Tantivy full-text + pgvector vectors + RRF fusion, streaming answers with inline citations;
- Ontology & cold start — ships with five ontology packs (schema.org, W3C Org, PROV-O, FOAF, IOF Core);
- Reasoning — ontology axioms compile into rules with forward-chaining (off by default, derived facts tagged with provenance and confidence);
- Conflict detection — three kinds of conflict, three sets of choices;
- Decision ledger — append-only record of who changed what and when;
- Ontology2SQL — ontology-driven text-to-SQL, state-of-the-art on the BIRD benchmark.
The part I respect most: it can run air-gapped
The model layer accepts any OpenAI-compatible endpoint (DeepSeek, Qwen, GLM, Ollama, vLLM), so the entire system can run fully offline, physically air-gapped.
For government, finance, and defense deployments, that's what "private" is actually supposed to mean — data, model, and inference all on hardware you control, with no network egress.
The honest caveat
Utopia is still at v0.1. The database schema evolves between versions and migrations roll forward only. Pin your version in production and back up the database along with the data directory.
If you're doing lightweight "chat with my PDFs," a plain RAG stack is fine. If you need a knowledge foundation your agents can trust, with a compliance trail — Utopia is worth a serious look.
I've localized the README and core docs to Chinese: https://github.com/yangshun2005/utopia-cn
If you find this project useful, a star on the original repo supports the author's ongoing maintenance.
Top comments (0)