Prediction Over Your Graph

A graph database answers what is connected. Aito answers what is probably true about the connections you have not seen yet: the missing edge, the label on an unlabelled node, the counterparty that does not fit. Every answer is a calibrated probability with the edges that drove it, and it reflects the next write with no retraining.

---

What it does

• Link prediction, including for nodes with no edges yet. Predict the missing endpoint of a relation. Condition on the node's attributes and a brand-new node gets an answer from what kind of node it is, not from a history it does not have.
• Node classification, with the reason. Label a node from its neighbourhood. $why names the edge conditions that moved the answer, so a person can check it and an agent can cite it.
• Suspicious edges. Predict who the counterparty should be. When the actual counterparty gets a low probability, the edge is flagged, with the fields that expected someone else.
• Grounding and corroboration for GraphRAG. Fetch an entity, its neighbourhood and the evidence behind it in one query. $why says which edges drove an answer, and $distinctLength counts how many independent sources back a claim.

Relationships are ordinary links. Aito follows them forward with dotted paths, as deep as your schema goes, reads them backwards with $refs, and combines both in one query, in the same store as your facts, text, and vectors.

Evidence read through a link, in Company AI, an open reference application, on its synthetic demo data. Whether a CTO is on file exists nowhere on the deal: the query goes forward to the account and back to its people, and $why names that condition and its lift.

---

Link prediction

Model your data as entities and edges, one row per subject, relation, target. Predicting target given subject and relation is link prediction:

POST /api/v2/_predict
{
  "from": "edges",
  "where": { "subject.role": "tech", "relation": "uses" },
  "predict": "target",
  "basedOn": ["kind", "industry"]
}

Conditioning on subject.role rather than a subject id is what lets a new node with no edges get a prediction. basedOn does the same on the other side: it ranks candidate targets by their own attributes, so a target that has never been linked is still a candidate.

---

Node classification

Classify an entity from its neighbourhood, here a segment for anyone who has engaged with something:

POST /api/v2/_predict
{
  "from": "entities",
  "where": { "$refs.edges.subject": { "$exists": { "relation": "engaged_with" } } },
  "predict": "segment",
  "select": ["$value", "$p", "$why"]
}

The reverse-linked edges act as evidence with no data copied onto the entity, and $why names the edge condition behind the answer.

The same in Company AI: what kind of company this is, judged only by who works there. Synthetic demo data.

---

A counterparty that does not fit

Fraud screening on a graph usually means searching for suspicious patterns. Aito asks a more direct question: given everything else about this edge, who would we expect on the other end?

POST /api/v2/_predict
{
  "from": "payments",
  "where": { "payer.segment": "retail", "category": "consulting" },
  "predict": "payee",
  "select": ["$value", "$p", "$why"]
}

If the payee actually on the payment sits far down that list with a low $p, the payment is flagged, and $why shows which fields pointed to someone else. It is a probability to threshold, not a rule to maintain.

---

Grounding for GraphRAG and agents

An agent answering from a knowledge graph needs two things beyond retrieval: which edges support the answer, and how many independent sources stand behind them. $why gives the first. $distinctLength gives the second, counting distinct sources rather than rows, so one source filing the same claim ten times reads as one corroboration, not ten:

POST /api/v2/_query
{
  "from": "claims",
  "select": [
    "id", "subject", "relation", "target",
    { "sources": { "$distinctLength": "$refs.evidence.claim.source" } }
  ]
}

Writing a fact is adding an edge, and retracting it is deleting one. The next query reflects it.

---

What it is not

Aito is a predictive database that reads links both ways, not a graph traversal engine. Any path you can name, you can query: forward links chain to any depth, and $refs reads a link backwards. What it does not do is search for paths whose length you do not know in advance, such as reachability, shortest paths, or everything below a node in a hierarchy. Today that is a bounded agent loop, one query per step. There are no graph algorithms such as PageRank or community detection either. If those are the core of your workload, run them in a graph database, fed from Aito over SQL. Where Aito fits is the other half: predicting the edge a graph database does not have yet, and saying why.

---

Get Started

Start for free → Model your entities and edges and predict over them.

Read the graph docs → The full operator surface: links, $refs, node classification, link prediction, and provenance.