ADR-005: Knowhere vs Eagle Query Boundary¶
Status: Accepted
Decision¶
- Eagle uses Knowhere for parse and
doc_navonly. - Query hot path: Eagle Milvus multi-collection + RRF — not Knowhere
RetrievalAgent/WorkflowOrchestrator. - Parent-document retrieval: two-stage on
eagle_textinKnowhereGraphRetriever. - Structural parse is Knowhere-only for every industry document: section tree, TOC, typed chunks, and hierarchical
pathmust be preserved from KnowhereParseResult. - Domain
CHUNK/PARSEhooks enrich only (metadata, soft labels, optional typed annotations). They must not discard or rewrite Knowhere skeleton fields (path, chunk body,doc_nav,section_summarynodes). - If a domain needs different split / hierarchy boundaries, change Knowhere parse (upstream / parse-sdk) and consume the new
ParseResultin Eagle — do not implement a from-scratch Eagle chunker.