Skip to content

ADR-005: Knowhere vs Eagle Query Boundary

Status: Accepted

Decision

  • Eagle uses Knowhere for parse and doc_nav only.
  • Query hot path: Eagle Milvus multi-collection + RRF — not Knowhere RetrievalAgent / WorkflowOrchestrator.
  • Parent-document retrieval: two-stage on eagle_text in KnowhereGraphRetriever.
  • Structural parse is Knowhere-only for every industry document: section tree, TOC, typed chunks, and hierarchical path must be preserved from Knowhere ParseResult.
  • Domain CHUNK / PARSE hooks enrich only (metadata, soft labels, optional typed annotations). They must not discard or rewrite Knowhere skeleton fields (path, chunk body, doc_nav, section_summary nodes).
  • If a domain needs different split / hierarchy boundaries, change Knowhere parse (upstream / parse-sdk) and consume the new ParseResult in Eagle — do not implement a from-scratch Eagle chunker.