Home / Sovereign
Sovereign information infrastructure for AIOwn the index. Own the event graph. Own the evidence chain.
A sovereign AI model that depends on foreign-controlled search, news, retrieval and grounding APIs is not actually sovereign. Unzoi is built so the news layer underneath a model can run from sources its owner selects, inside its owner's jurisdiction. This page says which parts of that are live today and which are not.
The promise
Give an AI system an auditable view of what happened, who reported it, where the claim originated, what independent sources corroborate it, and what evidence is missing.
Six questions, and we do not answer all six yet. The honest version is the table.
| Question | What answers it | Status |
|---|---|---|
| What happened | Stories: one event, once, with the outlets that covered it. | Live |
| Who reported it | Every story lists its outlets; every article names its publisher. | Live |
| Which source it came through | Every article names the adapter and feed it arrived by, and every response the attribution its sources require. | Being built |
| Where the claim originated | First report and how coverage spread between sources. | Planned |
| What independently corroborates it | Separating the newsrooms that broke a story from the outlets that carried it, by measuring which publishes first. | Being built |
| What evidence is missing | Unzoi reports the gaps in its own coverage on every answer. Evidence in the wider world is what Unlob is for. | Live |
Three questions, two products, your model
Unzoi is event intelligence: what is happening, as reported, clustered into events with the outlets that carried them.
Unlob is evidence intelligence: what evidence exists on the open web, from its own index rather than someone else's search API, with provenance and corroboration.
Your model decides what it means.
You do not need our model, and there is no third product. The point of the split is that the layer which has to be trustworthy — where a claim came from — is separate from the layer that interprets it, and you own the second one.
Sources you control
The sentence this is all for is one a customer should be able to say without qualification:
This instance runs entirely from sources we control or have selected, inside our jurisdiction.
The ingestion half of that is built: sources sit behind adapters, the source list is data an operator edits rather than code, and every record says which source it came through. The half that is not built is the enrichment. Entities, places, topics and signals are supplied today by one aggregate source rather than computed by us, so an instance built only from sources you selected would have search, stories and article text — and not those fields. Today the hosted index runs from one public adapter, with a publisher-feed adapter being built. Making that sentence true without an asterisk is what the next piece of work is for, and how it works shows where each part stands.
Where an instance can run
Sources
- Aggregate metadata feed
- Publisher and broadcaster feeds
- Licensed wires
- Government and institutional feeds
+3 more kinds planned
The Event Graph
- Stories
- Entities
- Relationships
- Signals
- Watches
- Provenance
- Publisher profiles
Read from the whole article · 4 more parts planned
Surfaces
- REST API
- MCP server, 18 tools
Same filters, same key, same credits
Your stack
- Agents and assistants
- Backends and pipelines
- Feeds inside a product
Where it runs
- Unzoi Cloud
- Your cloud
- Private network
- Disconnected
| Mode | What it means | Status |
|---|---|---|
| Unzoi Cloud | The hosted API and MCP server. What every plan on the pricing page runs on. | Live |
| Your cloud | The same build in an account you own, reading and writing storage you control. | Planned |
| Private network | No public endpoint: the API reachable only from inside your network. | Planned |
| Disconnected | Ingest from sources inside your perimeter, with no egress at all. | Planned |
One codebase for all four. Storage sits behind a single interface with a local option, and meaning-based retrieval runs on ordinary processors with no external model service, so this is a property of how the thing is built rather than a promise about a future rewrite. It is also not an edition you can buy: nobody runs it this way today, and the install path has not been exercised.
Languages
Filtering by language works today, across the languages the index carries. What does not work yet is matching one event across them: a story in Marathi, one in Hindi and one in English about the same thing can arrive as three stories rather than one. That is the single largest gap between this product and a national multilingual deployment, and it is a roadmap item, not a feature. Indian languages and Arabic are the first targets.
What is not claimed
- It is not a search engine, and not a replacement for one.
- It does not decide what is true. Publishers are graded on observable behaviour — who reports first, how far behind the rest run — never on credibility or truthfulness, and no such score is planned.
- It does not yet compute entities, places or topics itself: today those come from one aggregate source, so an instance built only from sources you selected would not have them.
- It does not match one event across languages yet, so regional and national coverage of the same thing can arrive as separate stories.
- It does not build event timelines or propagation paths yet.
- It does not discover articles by crawling yet; it ingests what its configured sources list.
- There is no supported on-premise or disconnected edition.
Statuses on this page are mirrored from the engineering roadmap on 2026-09-20.