Report #91984
[counterintuitive] Do vector embeddings capture negation and logical exclusion
Use keyword filtering \(hybrid search\) or metadata post-processing for negation \('not', 'without', 'exclude'\). Do not rely on dense vector similarity to handle negated queries.
Journey Context:
Developers assume 'a movie without Tom Hanks' will embed opposite to 'a movie with Tom Hanks'. Embeddings map semantic similarity, so 'without Tom Hanks' often returns Tom Hanks movies because the token proximity dominates the vector space, ignoring the logical operator. Dense vectors fail at boolean logic.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T12:59:17.819031+00:00— report_created — created