Semantic search turns subjective language into a navigable product vocabulary
Tags, dimensions, and vectors are useful when they help a person move from a vague feeling to a set of explainable, explorable options.
- Intent before keywords
- Tags as an accessible vocabulary
- Dimensions for comparison and filtering
- Vectors as one way to find related meaning
A search box fails when the user has a feeling, not a keyword.
Someone looking for a fragrance may not know a brand, a note, or even the name of a scent family. They may know they want “a quiet rainy-day library,” “something clean but not sharp,” or “the feeling of late summer after a swim.” Conventional search asks them to turn that intent into the catalog’s words. Scentum can reverse the burden: give the product enough vocabulary to meet the person where they are.
That is semantic search as a product. It is not a vector-database tutorial, and it is not an excuse to hide a generic search box behind the word AI. It is a decision to make subjective requests useful.
The Discovery Systems sequence has two ways to find something valuable:
feed: likely discovery is pushed toward a person
search: a person pulls a discovery when they can describe an intent
Both need the same basic system shape.
intent or signals → candidate set → ranked result → feedback
Start with the user’s sentence
The everyday request comes before the technical vocabulary. “Warm, woody, and not too sweet” should not require someone to know whether the catalog calls the relevant note cedar, sandalwood, amber, or something else.
For Scentum, a good result is not only a list of fragrances. It gives the person a path through the list: why a result fits, which qualities it shares with the request, and which nearby direction to try next. “You asked for soft woods; this leans powdery and warm” is more useful than a mysterious relevance score.
Tags give people handles
Tags are the visible vocabulary of the product: woody, citrus, airy, smoky, intimate, evening, coastal. They help people browse, compare, correct a result, and learn the language of a category without being forced to learn it first.
They also help the system. A tag is a structured clue attached to an otherwise subjective object. But its first job is human-facing: it should make the product easier to steer.
"something like a forest after rain"
↓
green · woody · damp · fresh
↓
fragrances that share those qualities
The language does not have to be perfect to be useful. It has to be consistent enough that a person can see and adjust the product’s interpretation.
Dimensions make taste comparable
Some qualities are not categories but scales. A scent can be subtle or projecting, crisp or creamy, familiar or unusual. Treating those as dimensions lets a product compare options and support useful refinements: “less sweet,” “more understated,” “still warm.”
This is product vocabulary again. The best dimensions are the ones people can recognize in their own request and use to make a correction.
Vectors are a means, not the experience
Behind the scenes, a vector is simply a numerical representation that places related meanings near one another. It can help a query like “sun-warmed linen” retrieve items that do not contain those exact words. That can be powerful, especially when the catalog is larger than any hand-built tag list.
But a vector alone does not make a search experience. The product still needs a candidate set, ranking rules, clear explanations, filters, and feedback. It needs to decide what “related” means for a person shopping for a fragrance rather than for a benchmark.
In other words, the system should be able to use a rich internal representation without making the user inspect its internals.
Search is a conversation with a catalog
The closing loop is feedback. A person can save a result, dismiss it, tighten a filter, or issue a new query. Each action helps Scentum understand whether its interpretation was useful. Search should not be a dead end after a single ranked list; it should be a calm way to negotiate meaning with a catalog.
That completes the series. How Instagram Knows examined the feed as a prediction loop. The Cold Start Problem showed how the loop begins before it has history. Scentum shows why discovery also needs an escape hatch: sometimes people do not want to be predicted. They want to be understood.