Use vector search when meaning matters
A customer may ask “Can I freeze my membership?” while the policy says “temporary suspension.” Vector search can connect those phrases. This is useful when wording varies but the underlying intent is close.
Use normal search when precision matters
- Exact IDs, order numbers, SKUs, and dates.
- Permission and tenant filters.
- Financial totals and transactions.
- Queries where matching a known phrase is more important than conceptual similarity.
The hybrid pattern
A support search might first filter to one customer’s account and current policies with SQL, then use vector similarity to find the most relevant passage. The vector score helps rank candidates; it should not be the permission system or the source of truth.
Questions to answer before buying infrastructure
- How many documents need semantic search?
- How often do they change?
- What metadata must always filter the result?
- What is the cost of a false match?
- Can a simple full-text or database search solve the first version?