Architecture

多模态检索模式，结合语义搜索、精确/关键词搜索和近期搜索并行运行。结果合并并重新排序为单一上下文集。召回率高得多，因为智能体可以找到模糊引用和精确实体。对于全面的知识检索至关重要。

完整 FlowZap 代码

User {
  n1: circle label="Ask mixed query"
  n2: rectangle label="See answer"
  n1.handle(right) -> Agent.n3.handle(left)
  Agent.n18.handle(right) -> n2.handle(left)
}

Agent {
  n3: rectangle label="Plan retrieval strategy"
  n4: rectangle label="Trigger semantic search"
  n5: rectangle label="Trigger keyword search"
  n6: rectangle label="Trigger recent-history search"
  n7: rectangle label="Merge and rerank"
  n8: rectangle label="Build prompt with hybrid context"
  n9: rectangle label="Call LLM"
  n18: rectangle label="Return answer"
  n3.handle(bottom) -> n4.handle(top) [label="Semantic"]
  n3.handle(right) -> n5.handle(left) [label="Keyword"]
  n3.handle(left) -> n6.handle(right) [label="Recent"]
  n7.handle(right) -> n8.handle(left)
  n8.handle(right) -> n9.handle(left)
  n9.handle(right) -> LLM.n19.handle(left)
}

Semantic {
  n10: rectangle label="Vector search"
  n11: rectangle label="Return semantic matches"
  Agent.n4.handle(right) -> n10.handle(left)
  n10.handle(right) -> n11.handle(left)
  n11.handle(bottom) -> Agent.n7.handle(top) [label="Semantic"]
}

Keyword {
  n12: rectangle label="Exact/ID search"
  n13: rectangle label="Return exact matches"
  Agent.n5.handle(right) -> n12.handle(left)
  n12.handle(right) -> n13.handle(left)
  n13.handle(bottom) -> Agent.n7.handle(top) [label="Keyword"]
}

Recent {
  n14: rectangle label="Scan recent messages"
  n15: rectangle label="Return recent matches"
  Agent.n6.handle(right) -> n14.handle(left)
  n14.handle(right) -> n15.handle(left)
  n15.handle(bottom) -> Agent.n7.handle(top) [label="Recent"]
}

LLM {
  n19: rectangle label="Reason over hybrid context"
  n19.handle(right) -> Agent.n18.handle(left)
}

相关模板