Semantic search

The failure it addresses#

Keyword search fails in a specific and frustrating way: you remember the substance of a document and not its wording. A query for "the lease that renews in the third quarter" finds nothing when the agreement says "term expires 30 September", even though it is exactly the right document.

Semantic search handles that case because the two phrasings are close in meaning even though they share no keywords.

Why it should not replace exact matching#

When a user searches for an invoice number, a document reference or a person's surname, they want exact matches and nothing else. Semantic matching on a precise identifier returns plausible near-misses, which is worse than returning nothing.

The right implementation returns both and presents them together, rather than making the user choose a retrieval algorithm.

Practical caveats#

Semantic results are harder to explain — a user cannot see why a document matched, which undermines trust when the match is poor. And the same permission filtering applies: a vector index must respect access rights at query time, or it becomes a route to content a user cannot open.

Where it is most valuable#

On large archives of narrative documents — correspondence, reports, meeting records — where the vocabulary varies between authors and no naming convention was ever enforced. On structured, well-classified content it adds less, because metadata already answers the questions people ask.