How nodes answer each other without sharing files

Picture a question that needs a colleague's expertise. One option is to have their whole filing cabinet wheeled over to your desk so you can read the source pages yourself. The other is to send the colleague the question and get back a short written brief. Both answer the question; only one moves the files. Kovalent's mesh supports both, and which one a node uses depends on its tier.

On the higher tiers (Team and Enterprise), a node that routes a query to the best-matched peer asks that peer to synthesize an answer rather than to hand over source text. The peer runs the search against its own store, composes a compressed answer under its own cloud identity, and returns only that answer. The requesting node then does the final assembly. This is the map-reduce shape in miniature: the peer maps its own matching passages down into a single brief, and the requester reduces that brief into the response it returns to you.

By default, the peer has that brief written by a cloud inference service it calls under its own credentials, so answers come back in seconds and the control plane never sees the request. Deployments that need to keep inference on the node itself can switch to sovereign mode, where a model inside the node writes the brief. Either way, only the finished brief crosses the mesh.

On the Community tier, a node instead pulls the matching raw chunks from its peer into its own context window and assembles the answer itself. It is simpler and works well within a single tenant's own mesh, where every node already belongs to the same trust boundary. When a node has no reachable peers at all, it does not fail; it answers from its own local store and records that it handled the query alone.

Whichever route runs, the node routes to a peer by comparing fingerprint similarity, so the query goes to the peer most likely to know the answer rather than to everyone at once. Every hop of this exchange emits a metadata trace for observability.

Synthesis trades detail for containment: a compressed brief is not the source passages, so a reader who needs the exact original wording is better served by a node that owns the document directly. Peer calls run under a timeout, so an unreachable or slow peer surfaces as a failed exchange rather than a hang. And the routing is only as good as the peers' published domains; a query far outside any peer's expertise still lands on the closest available match.