In this piece
Utopia Review: Can Your Knowledge Graph Rewind Business History?
Your assistant answers who owns a customer account today. Can it explain who owned it when an exception was approved, using only evidence available at the time? That is the buying question behind Utopia. A polished chat interface does not answer it by itself.
Our recommendation: evaluate Utopia for a bounded workflow where late corrections and historical evidence affect decisions. Keep a simpler search system if people only need the current handbook. This is a documentation review dated 6 September 2026, pinned to commit 9a3ab3373244, not a deployment benchmark or a client implementation.
What is Utopia's enterprise world model?
Utopia is DeepLethe's self-hosted, Apache-2.0 knowledge application. Its official repository describes a Rust and PostgreSQL stack, document search, an ontology and a bitemporal graph. In this context, “world model” means structured enterprise knowledge that changes over time. It does not establish that the system can predict business outcomes.
The useful distinction is between finding a sentence and maintaining the business assertion extracted from it. A source might say a manager changed roles. A governed system must resolve the person, role, effective date and conflicting evidence before that assertion becomes a dependable answer.
Why do historical answers need two clocks?
Valid time asks when a fact applied in the business. Record time asks when the system knew that version. A late correction changes today's understanding of the past without changing what was known on the earlier date. Both dates must be explicit when evaluating an answer.
Consider this fictional account-ownership fixture. These are expected outcomes for a pilot, not measured Utopia results. Treat each interval's end as exclusive.
| Evidence arrives | What it says | Question and expected answer |
|---|---|---|
| 1 March | Ada owns account Orion from 1 March. | Who owned Orion on 12 March, as known on 12 March? Ada. |
| 20 March | A signed correction says Ben took over on 10 March. | Who owned Orion on 12 March, using today's corrected evidence? Ben. |
| After the correction | The earlier belief remains historical evidence. | Replay what the system knew on 12 March: Ada, with the earlier source. |
Utopia's two-clock design record separates at for business time from as_of for record time. It describes historical graph reads and vector retrieval, but says full-text recall remains current-only and the second-clock UI control is unresolved. A timeline demonstration therefore does not prove complete historical retrieval. Test the API, retrieval path and interface you will actually deploy.
Does every fact have an exact start and end date?
No. Real evidence often omits dates. Utopia's unknown-date design distinguishes an unknown ending from an assertion that still holds. Missing dates must not silently turn into “always true.” Its timestamp design describes sub-day precision with explicit time zones, although the pinned README still lists finer precision on the roadmap. Verify the chosen release instead of treating a roadmap checkbox as a capability test.
Add three cases to your evaluation: a document with no start date, a statement that a role ended without saying when, and two handovers on the same afternoon. Require the answer to expose uncertainty. An upload date can be an evidence anchor; it is not automatically the date a business event happened.
Does a source citation prove a fact is correct?
A citation provides a route to evidence. It does not prove that the extracted assertion, date or inference is right. A sentence can describe a proposal, a rejected option or somebody else's allegation. Review whether it actually supports the answer.
The provenance export design documents evidence quotes and document lineage for asserted facts, and rule-and-premise chains for derived facts. That is more precise than saying every conclusion appears verbatim in one sentence. It also says the RDF export is not a restorable backup.
For each accepted answer, preserve the fact identifier, requested dates, document version, supporting quote and any derivation chain. Have a reviewer check the source without seeing the model's explanation first. Otherwise a persuasive explanation can hide a weak citation.
What should a company verify before self-hosting Utopia?
The pinned project is an early v0.1 system. Its README lists enterprise work including OIDC SSO and describes forward-only migrations. Offline operation requires local model endpoints as well as local application storage. Hosting the web app yourself does not keep prompts local if it calls an external model.
Read the security guidance before exposing a pilot. It calls for least-privilege source credentials, controlled workspace grants and preserving the encryption key with the data directory. Our evaluation recommendation is to test restoration, revoked access and historical-source permissions before importing sensitive records.
- Release: record the image digest, schema version, model and embedding configuration. Re-run historical fixtures after upgrades.
- Access: create users with different source entitlements. A historical query must not reveal a document a user cannot read now.
- Corrections: nominate a business reviewer who can settle competing assertions and explain why.
- Recovery: restore the database, files and required keys into an isolated environment, then repeat the same questions.
- Exit: inspect exported history and provenance separately from backup restoration. Budget migration work if another system must consume them.
For the broader access-control design, use our permissions-first RAG guide. For a technical scoping engagement, see RAG and AI architecture services.
When is Utopia a better candidate than a simpler approach?
Temporal knowledge is not unique to Utopia. Graphiti is a framework for temporal context graphs, while SQL Server temporal tables preserve row history. Utopia's appeal is the assembled knowledge application. The following choices are our architectural judgment, not a comparative benchmark.
| Main requirement | Starting point | Decisive pilot question |
|---|---|---|
| Current handbook answers | Existing search or permission-aware RAG | Is history actually needed to answer correctly? |
| Structured ownership records in an existing database | Database history plus business-effective dates | Can SQL answer this without extracting facts from prose? |
| A custom agent that needs temporal memory | A framework such as Graphiti | Does the team want to own the application layer? |
| Changing facts spread across documents, with human review | A controlled Utopia pilot | Can it reconstruct both business state and earlier understanding? |
Use our graph-engineering guide for the general graph-versus-retrieval decision and OpenKB review for the separate compiled-wiki approach. If the decision is whether to adapt a product or commission software, start with build-versus-buy guide.
A ten-day Utopia evaluation plan with acceptance criteria
This is a proposed pilot scope. Adapt the numbers to your risk and volume; they are not vendor performance claims.
- Days 1–2: choose one workflow, such as account handovers. Collect 30 approved document versions and write 40 questions: 10 current, 10 historical, 10 late-correction and 10 unknown-date or access-denied cases.
- Days 3–4: define the entities, relationships and date rules. Label expected answers with the process owner. Keep a held-out subset out of ontology tuning.
- Days 5–6: run Utopia and your current retrieval baseline against the same corpus and questions. Record citation support, both clock values, abstentions, latency and model usage.
- Days 7–8: introduce the late correction, revoke a user's access, replace a source and restore a backup. Re-run the held-out questions after every change.
- Days 9–10: review errors with the business owner. Decide whether the temporal benefit justifies curation and operations.
Suggested gates: every answer used for a decision has a checkable source; all designated late-correction fixtures select the expected version; no restricted evidence leaks in the access tests; unknown dates are not invented; and the restored system reproduces the fixture results. Set workload-specific latency and quality thresholds before testing. Passing a small set supports a next pilot phase, not a blanket production guarantee.
What does a Utopia pilot cost?
The license is only one line in the budget. Estimate document parsing and extraction, embedding and chat inference, database and file storage, reviewer time, integration, monitoring and recovery. The cost driver most easily missed is maintaining the ontology and resolving disputed facts.
Monthly operating cost = infrastructure + model usage + review hours × internal hourly cost + maintenance. Divide that total by accepted, source-supported answers for a useful unit cost. Compare it with time spent reconstructing history manually. Do not use a generic per-query price without the corpus size, change rate and review workload.
Our TwinSoft AI case study illustrates adjacent AI integration work, not a Utopia deployment. If historical answers matter to your team, scope a temporal-knowledge pilot with Wavect. Bring two conflicting document versions and one decision you need to reconstruct. Those inputs make the architecture discussion concrete.
Utopia evaluation FAQ
Is Utopia a replacement for RAG?
It is a candidate for knowledge workflows where history and relationships matter. Ordinary retrieval can remain sufficient for current-document questions. Compare both on the same labelled workload.
Can I rewind the graph and trust the answer?
Only after verifying which clock is selected and which retrieval components honour it. Historical graph state, historical vector search and full-text history are separate acceptance tests.
Should we deploy it across the company immediately?
Start with a restricted pilot and a named reviewer. Expand when correction handling, access controls, restoration and operating cost meet agreed criteria.