The framing I find most useful here is that vectorless approaches do not remove chunking, they move the decision from ingest time to query time. A table of contents or a wiki page is still a boundary someone drew, the difference is that navigation can choose to descend or not, which is exactly why it wins on long structured documents where the answer depends on which section you are in. The cost point in section 9.1 is the honest catch: navigation is several model calls per query with latency you cannot cache away, versus one vector lookup measured in milliseconds. The hybrid section is where I would expect most production systems to land, and the routing signal that has worked for me is document structure rather than query type - navigate when the corpus has real hierarchy, embed when it is a pile of independent passages.
The framing I find most useful here is that vectorless approaches do not remove chunking, they move the decision from ingest time to query time. A table of contents or a wiki page is still a boundary someone drew, the difference is that navigation can choose to descend or not, which is exactly why it wins on long structured documents where the answer depends on which section you are in. The cost point in section 9.1 is the honest catch: navigation is several model calls per query with latency you cannot cache away, versus one vector lookup measured in milliseconds. The hybrid section is where I would expect most production systems to land, and the routing signal that has worked for me is document structure rather than query type - navigate when the corpus has real hierarchy, embed when it is a pile of independent passages.