DevOps

Un pattern de routage qui place un routeur sémantique devant les outils MCP pour que le LLM ne voie que le sous-ensemble dont il a besoin. Utilise des embeddings vectoriels et la similarité cosinus pour matcher l'intention utilisateur aux outils dynamiquement. Réduction jusqu'à 96% des tokens d'entrée avec les grands catalogues d'outils.

Code FlowZap complet

Host {
  n1: circle label="User sends prompt"
  n2: rectangle label="Agent extracts intent"
  n3: rectangle label="Send intent to router"
  n4: rectangle label="Receive routed result"
  n5: rectangle label="Agent responds to user"
  n1.handle(right) -> n2.handle(left)
  n2.handle(right) -> n3.handle(left)
  n3.handle(bottom) -> Router.n6.handle(top) [label="Intent + tool request"]
  n4.handle(right) -> n5.handle(left)
}

Router {
  n6: rectangle label="Receive intent"
  n7: rectangle label="Semantic match via embeddings"
  n8: diamond label="Which MCP server?"
  n9: rectangle label="Forward to Server A"
  n10: rectangle label="Forward to Server B"
  n11: rectangle label="Normalize and return result"
  n6.handle(right) -> n7.handle(left)
  n7.handle(right) -> n8.handle(left)
  n8.handle(bottom) -> n9.handle(top) [label="Route A"]
  n8.handle(right) -> n10.handle(left) [label="Route B"]
  n9.handle(bottom) -> ServerA.n12.handle(top) [label="Call Server A"]
  n10.handle(bottom) -> ServerB.n14.handle(top) [label="Call Server B"]
  n11.handle(top) -> Host.n4.handle(bottom) [label="Final result"]
}

ServerA {
  n12: rectangle label="Execute tool A"
  n13: rectangle label="Return A result"
  n12.handle(right) -> n13.handle(left)
  n13.handle(top) -> Router.n11.handle(bottom) [label="Result A"]
}

ServerB {
  n14: rectangle label="Execute tool B"
  n15: rectangle label="Return B result"
  n14.handle(right) -> n15.handle(left)
  n15.handle(top) -> Router.n11.handle(left) [label="Result B"]
}

Modèles associés