schema · growing
Metadata filtering
"2021 risk factors" is not a meaning, it is a filter. Raw semantic search confuses the two.
Metadata is context you can inject into every chunk: year, document title, page, section, owner. Raw semantic search confuses these at low precision.
Rendering diagram…
The 2021 inside the question is not a meaning, it is a filter. Vector
similarity sees the 2020 and the 2021 document as nearly identical.
Rendering diagram…
What to store
Structuralyear, page, section, doc type
Derivedchunk summary, summary of neighbours
Reverse HyDEwhich questions this chunk answers