Vector search needed product questions in 2023

Vector search answers a similarity question. The product still has to define what similarity means.

For support, it may mean the same customer problem. For legal discovery, the same clause. For code search, the same behavior. For incidents, the same failure pattern. Those are different products.

Search and answering also have different failure modes. Search exposes documents. Answering synthesizes a claim that can hide a bad retrieval result. The system should know whether it is building search, Q&A, recommendations, research, triage, or memory.

Similarity is only one signal. Combine it with authority, freshness, permissions, source quality, role, and workflow state:

score =
  semantic_similarity * 0.55 +
  source_authority * 0.20 +
  freshness * 0.15 +
  user_affinity * 0.10

Empty results are part of quality. A nearest neighbor is not necessarily an answer:

No strong match found.

Try searching by customer name, invoice ID, or policy section.

Filter permissions before a result reaches the model. Once retrieval feeds an answer, citations are part of the interface:

Source: Refund policy v3, "regional exceptions", updated 2023-09-14

Test near misses: account deletion versus workspace deactivation, trial refund versus renewal refund, API-key rotation versus session revocation, production outage versus staging failure.

The 2023 vector-search wave made retrieval accessible. The product questions decided whether it was useful.