A keyword search matches letters. A vector search matches meaning. They
sound similar; they return almost opposite things. Type a word and watch the gap — then keep
scrolling to see why it happens.
try:
·
·
·
·
·
loading embeddings…
two searches, one query
keyword search — shared letters (trigram)
vector search — shared meaning (cosine)
Words in both columns are the easy cases —
where spelling and meaning happen to agree. Click any word to search it.
the same hits, on a map of meaning
Position ≈ meaning (a 2-D PCA shadow — indicative, not exact). The
keyword hits scatter across the whole space — words that share letters can
mean anything — while the vector hits collapse into one neighborhood.
That scatter is the lexical gap, made spatial. Click a labeled dot to search it.
meaning is a navigable web
The query and its nearest meanings, as a graph. Orange = expanded;
blue = a leaf you can grow — click any leaf to sprout its own neighbors and
wander concept-space (click an expanded node to re-center). Keyword search has no such web: its
links are spelling accidents that dead-end.
The vocabulary problem. Two people pick the same word for the same thing less than
20% of the time
(Furnas et al. 1987),
so a keyword index — matching surface form — misses most of what a searcher means, while returning
look-alikes that mean nothing related. Vector search closes that lexical gap by comparing
meaning. (This is a word-level stand-in for document/passage retrieval, but the gap is the same.)