"The payment succeeded" and "the payment did not succeed" are lexically almost identical and often embed close together. If your system routes on similarity alone, it will confidently retrieve the wrong one.

Where this bites

• Negation, which embeddings handle poorly across most models.
• Numbers and dates, where "before 2020" and "after 2020" look similar and mean the opposite.
• Named entities that differ by one token: two products, two people, two account numbers.

What to do about it

Filter structurally before you rank semantically. If the question mentions a date range, a product or an account, extract it and use it as a hard filter — then let vector similarity order what remains. This is the same insight as hybrid search stated differently: embeddings are excellent at "roughly about the same thing" and unreliable at "exactly this thing", so use each for what it is good at.