Identity of an embedding space: vectors are comparable only when all three
fields match. Recorded on a knowledge base at ingest and compared with the
active query backend before every retrieval.
A knowledge base was indexed in a different embedding space than the one
the active backend produces. Searching it would compare incompatible
vectors and silently return nothing, so callers must surface this instead.
How many chunks to retrieve and how similar they must be, for an agent that
searches configs together. Each KB falls back to the defaults for what it
leaves unset. Across KBs the most permissive setting wins (lowest
min_score, highest top_k), so a KB tuned for a lower-scoring embedding
space is not filtered out by another KB’s stricter setting.