Report #73508
[counterintuitive] Vector embeddings capture semantic negation for filtering
Do not rely on vector similarity to filter by negation \(e.g., 'not bad', 'without X'\). Use keyword filtering \(hybrid search\) or metadata filtering for exclusion.
Journey Context:
Developers assume 'good' and 'not good' will have distant embeddings because they mean opposite things. In reality, embedding models map 'not good' very close to 'good' because they share almost all contextual tokens and often co-occur. Cosine similarity cannot reliably represent logical negation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T05:58:37.820620+00:00— report_created — created