A graph-and-vector architecture for exploratory retrieval: finding not only what is near a query, but which inspectable relation might connect otherwise distant regions of a semantic space.
When a model searches for what is close to a query, it often finds synonyms, familiar associations, and more examples of the same idea. That is useful for lookup, but limited for exploration.
When asked to go further, a model may combine nearby words into a fluent explanation simply because they frequently occur together. The result can look grounded while containing little more than semantic proximity and verbal completion.
Exploration needs a richer coordinate system: one that can locate not only similar concepts, but also relations, tensions, contradictions, and intermediate structures. BaryGraph makes those relationships retrievable. A result may be a bridge, a contradiction, a higher-order pattern, or only a single useful term. It does not require every result to become a full answer.
Exploratory questions do not always need a longer answer. They often need a better next coordinate: a distinction, bridge, or relationship that changes what the question could mean.
Two concepts become the children of a BaryEdge, which receives its own stored vector.
BaryEdges become children of higher-order MetaBary triads, recursing upward into a relational forest.
Synonyms show proximity. Opposing terms create direction. BaryGraph preserves some apparent mismatches as navigational tension worth inspecting.
Nodes, edges, contradictions, and triads live in one index, reachable by ordinary nearest-neighbor search.
The name borrows from the barycenter in physics: the shared center generated by two bodies is treated as a real location with properties of its own. BaryGraph applies that intuition to semantic relations. It gives models a space containing not only points of similarity, but relational directions and opposing terms.
At the leaf level, a simplified construction is:
At higher levels, construction is algebra on vectors already stored in the graph — no new embedding call is required for each recursive step.
| Flat vector retrieval | BaryGraph retrieval |
|---|---|
| Finds text or concepts near a query | Finds concepts, relations, contradictions, and higher-order patterns |
| Primarily rewards semantic similarity | Represents similarity, distance, tension, and opposition |
| Often returns synonyms or familiar neighbors | Can surface a non-synonymic bridge between distant terms |
| May continue toward a complete answer | Can stop at a term, relation, or short hypothesis |
| Similarity is mostly attached to documents | Meaning can be attached to the relationship between documents |
| Strong for direct relevance and lookup | Designed to surface bridges, tensions, and intermediate structure |
| Often preserves the current framing | Can introduce tension that challenges the current framing |
Four independent language models probed the Kaikki proof of concept with open-ended questions and described what they found. These examples illustrate an interaction pattern, not benchmark performance: the system may return a useful bridge, a weak edge, or only a promising term.
“It behaves more like an associative atlas than a question-answering system. It found a genuine invariant across unrelated domains.”
“The graph refused to let boredom be an absence. It paired it with
overstimulationon the same axis. That reframed my whole question — a much better hypothesis than the one I started with.”
“A weak lexical bridge —
turn on— prompted a stronger functional reading: redundancy is often latent until activated.”
“A de-cliché device.”
Models are good at producing the first plausible association. BaryGraph introduces friction: a distant term, a contradictory relation, or a bridge that is not immediately explainable. It preserves that tension instead of smoothing it into another familiar continuation. The goal is not to make every answer strange, but to prevent familiarity from being mistaken for insight.
A probe around boredom, ritual and counting returned mensural, metrical,
cadenced, foot-tapping, time step. The graph supplied the
unexpected hinge: meter. The possible coordinate was
undifferentiated time → imposed measure → ritual and number, not a synonym for any one of them.
Ecological overlap, duplicate components, repeated message structure and removable legal surplus converged on one distinction: useful redundancy = multiplicity + failure independence.
Mycelial networks and internet architecture surfaced anastomosis beside interconnect
terms — precise vocabulary for asking which topological properties are substrate-independent.
An inactigram placed against memory / blank looked malformed as a statement, but sharp as a coordinate: recording inactivity means it is no longer blank. The contradiction opened an inspectable direction without claiming a conclusion.
The Kaikki English proof of concept, built and queried locally.
Ask an open question and inspect nodes, BaryEdges, MetaBary structures, contradictions, parent chains, and weak bridges. Treat the result as a possible direction, not an answer.
Ask what the relation changes. If it reveals a useful distinction or intermediate pattern, investigate it. If it does not survive inspection, reject it.
A useful result may be one term, one relation, or a short hypothesis. Do not expand it merely to satisfy a target answer length.
Use sources, domain knowledge, experiments, formal models, or human judgment to decide whether the coordinate deserves development.
A retrieved bridge is not asserted to be true, causal, or explanatory. A BaryGraph result is a prompt for investigation, not a proof. Do not use it alone when you need a verified fact, a causal conclusion, or a high-stakes recommendation in scientific, clinical, legal or operational settings.