Two customers open the same catalogue. One knows the part number. The other knows the problem they need to solve.
A useful search box should serve both. Adding AI does not remove the need to distinguish those intentions.
The following garden-supply catalogue and queries are fictional. They provide a simple way to discuss search design without pretending that a particular engine has been tested or that one approach wins everywhere.
The customer who already knows
Search: “replacement seal GS-240”
This customer is not asking for inspiration. They want a particular component, and the code carries information that should survive the search process.
Returning a beautifully written page about watering systems would be unhelpful if the required seal is available but buried below it. A near match can also be worse than no result when a customer mistakes a similar-looking part for the correct one.
Keyword retrieval gives the system a way to look for the actual terms in the request. Microsoft's hybrid search overview specifically identifies product codes, specialized jargon, dates and names as cases where keyword search can be useful for precise matching.
That does not make every keyword configuration equally good. A catalogue still needs consistent identifiers and a considered treatment of punctuation, alternate names and obsolete codes. If one supplier writes a code with a dash and another omits it, the product team should decide whether those strings identify the same item. An embedding model should not be left to invent that business rule.
For this journey, success means finding the correct item or clearly explaining that it is unavailable. A plausible substitute belongs in a visibly separate recommendation, with compatibility supported by the catalogue's own data.
The customer who knows the situation
Search: “something to water patio pots while I am away”
This request describes an outcome. The relevant product page might discuss a reservoir, a timer or a self-watering container without using the customer's exact wording.
Vector search represents content and queries numerically so it can retrieve information by conceptual similarity. That gives it another route to a relevant result when the vocabulary differs. It does not establish that a product satisfies every practical requirement.
For example, a suggested watering kit might require an outdoor tap. A product can be related to the query and still be unsuitable for the customer's patio. Availability, required connections and product specifications need to remain visible; a confident summary should not hide the unanswered questions.
Here, success is a useful set of options and enough information to narrow them. The product team may need to ask a follow-up question rather than immediately declare one item the answer.
Where hybrid search fits
Hybrid search combines keyword and vector retrieval instead of requiring the whole catalogue to choose one. Weaviate's documentation describes running both and combining their scores into a ranked result list. Azure AI Search also documents parallel full-text and vector queries, using reciprocal rank fusion to merge the results.
Combining results introduces choices. The balance between retrieval methods, the indexed fields and any additional ranking can change what a customer sees. Hybrid is an approach to evaluate, not a label that proves the correct product will come first.
The two fictional searches show why a single overall relevance score can conceal a poor experience. An improvement for broad discovery does not excuse a regression for customers entering a precise code. Review these journeys separately before deciding whether the combined experience is better.
A practical comparison can use representative customer queries, expected useful results and the same catalogue snapshot. Record where the required item appears, whether inappropriate alternatives are presented as exact matches, and whether the user can finish the task. These are proposed evaluation criteria, not measurements performed for this article.
The search interface matters too. A clear exact-match result, a separate related-products area and visible specifications can communicate distinctions that a blended ranking alone cannot. If an AI answer is added above the results, it should preserve those distinctions rather than turn every retrieved item into a recommendation.
The customer carrying a product code and the customer describing a problem are both expressing intent. Good AI search makes room for both, with a clear account of what the system found and what it still does not know.