Your documents know facts. A graph knows how they connect.
Entities, typed relationships and the ability to walk between them — so an agent can answer questions no search over text can reach. And the graph builds itself, inside a ceiling you set.
One template, four levels, stopped at the ceiling
Illustrative dataSome questions have no passage to find.
Grounded retrieval is excellent at finding the paragraph that answers you. It fails when the answer is an absence, or spans four documents that never mention each other. Every function has a handful of those, usually the ones that matter.
“Which controls have no evidence?”
- …evidence of review is retained for twelve months…
- …supporting evidence should be attached to each…
- …the evidence register is maintained quarterly…
Returns passages containing the word “evidence”. Nothing in any document states which controls lack it — the answer was never written down.
38 controlsevery entity of type Control← satisfiesfollow incoming evidence edges6 with nonethe empty set is the answer
Walks every Control along its satisfies edge and returns the ones with none. The answer is computed, not retrieved.
- an absenceWhich invoices have no matching receipt?
- an absenceWhich contract obligations have no owner?
- a pathWhich live services depend on a supplier nobody has assessed?
- a pathWhich finished products trace back to a single plant?
Two shapes. One answer is an empty set; the other is a route between things that never mention each other. Both are the same walk over typed edges, and the walk does not change when the domain does.
Every node is typed. Every edge is named.
An entity is not a blob of text with a label on it. It carries properties, tags and translations, and remembers the documents it came from. Its relationships point in both directions, so a traversal knows what it is following.
Applied control A.5.1
ControlcontrolinstantiatesPolicies for information securityControlRef
satisfiesInformation security policy — approvedEvidencesatisfiesPolicy acknowledgement recordsEvidence
Direction matters. What a control instantiates and what satisfies it are different questions, and the graph answers both.
Describe the shape. It builds the rest.
The examples below leave compliance for geography, because that is where this is easiest to see: a template describes one relationship, and the graph applies it outward until it hits a limit you set.
“Expand India to all states, then cities, and generate four images for each city”
The planner turns that into an execution plan you approve before anything runs.
- Source type
- location:country
- Target type
- location:state
- Relationship
- located_in
- Max per parent
- 50
List all states and provinces of {{parent.canonical_name}}.{{parent.canonical_name}} resolves per node, so one template covers every country in the graph.
- Max entities10,000
- Max depth4 levels
- Max cost$100
An autonomous process that spends money needs a ceiling more than it needs a start button. All three are set before the run begins.
A graph is also a production plan.
Select a type, add modifiers, and generate for every node at once, with the bill estimated before you commit. The examples below come from brand and content work, where one graph drives thousands of assets.
- Style + LocationArt Deco + New York
- Event + LocationDiwali + Tamil Nadu
- Hazard + SiteConfined space + Plant 4
Cross two entity types and generate what sits at the intersection.
- Entities
- 340
- Variations each
- 4
- Cost per image
- $0.167
Throughput and price are shown against your own deployment capacity, so the estimate is yours rather than a list price.
An answerable map of how your organisation actually works.
Nobody drew this. It accumulated: every document you indexed, every entity the expansion found, every relationship it named on the way. The graph is your organisation's shape: what you work on, how it connects, where the gaps are. Unlike the org chart or the wiki, it answers questions.
The first week, people ask it things they already knew. The second week, they ask it things nobody could answer before.
A missing owner, a missing receipt, a missing runbook: the graph hands those back as answers, before someone else finds them.
The entities the next question needs are already there, so the next question is a traversal rather than a project.
The graph, by the spec.
- Entities
- Typed, with properties, tags and assets
- Relationships
- Named and directional, in and out
- Exploration
- Search, filter and traverse by type
- Expansion
- Templates applied autonomously
- Planning
- Natural language to an approved plan
- Limits
- Max entities · depth · cost per session
- Generation
- Images and video per entity, in batches
- Languages
- 120+ supported
- Provenance
- Every entity stamped with its source
- Compliance
- SOC 2 Type II · HIPAA · GDPR · ISO 27001
Where every node came from.
A graph that writes itself is only trustworthy if you can see it doing so. Every entity and relationship is recorded as it appears — what was created, by which route, and when.
- entity createdvia expansion09:41Control · Access control policy
- relationship addedvia expansion09:38satisfies · Access review Q1
- relationship addedvia api09:24supplies · Assembly line 3
- entity createdvia import09:12Supplier · Northgate Components
And it never leaves your network
The graph is built, stored and traversed inside your own infrastructure, like everything else on the platform. No entity, no edge and no expansion run leaves the network it was created on.
- Stored on your infrastructure
- Scoped by group and role
- Every change audited
- Air-gapped deployment available
A map of how your organisation works is not a thing to hand to a vendor.
See Exemplary AI in action
Book a demo and we'll show you purpose-built agents grounded in your own knowledge — deployed on your infrastructure.