Technical specifications
Feature under development
Technical specification detection is under development and not yet generally available. It is enabled per solution on request, and the behavior described here may change in future releases.
A technical specification is the part of a product name that states a measurement, a rating, or a standard: Ø360x76cm, 7W, IP65, 5W30, 2 pcs. Technical specification detection lets Ecommerce Search read that text as one unit with a meaning, instead of as ordinary words.
Why ordinary word matching is not enough
A shopper rarely writes a specification the way the product data does, and small differences in symbols and spacing then decide whether a product is found:
| Search phrase | Without detection | With detection |
|---|---|---|
pool 360 |
No match for Pool Ø360x76 cm, because the 360 sits inside a longer word |
Finds it, because the numbers inside a specification can also be matched on their own |
GU 10 |
Two unrelated words, so any product with a 10 in its name competes |
Read as the lamp socket standard GU10 |
27 W |
Matches Combi oven 27 L 900 W, which is neither 27 W nor relevant |
Read as a power rating of 27 W, so that product is not a match |
5W30 |
Split into a power rating and a number | Read as the oil viscosity grade it is |
The pattern is the same in all four cases. The shopper and the product data mean the same thing, and only the writing differs.
How a search phrase is handled
When a search phrase contains a specification, Ecommerce Search reads the phrase in two parts:
- The specification becomes a requirement. A product has to carry that specification to be a match.
- The remaining words are matched and scored the way they always were.
Both parts have to be satisfied, so pool Ø360x76cm returns the products that are pools and carry those measurements. A specification narrows the result set, it never widens it.
Specifications are read out of free text. Filters remain the way to search on structured product data.
When a specification is read wrongly
Reading meaning out of free text is never perfect, so the feature is built so that a wrong reading cannot lose a product. When no product carries the detected specification, its original text is matched as ordinary words instead, exactly as the shopper typed it. That gives the same result as before the feature existed.
A phrase that holds no specification at all is unaffected, and search behaves exactly as it does today.
Effect on the order of results
A product that matches a specification more completely ranks above one that matches only part of it. A pool that is Ø360x76cm therefore ranks above a pool that merely has the number 360 somewhere in its name.
How much a specification can lift a product follows the relevance already configured for the product name, so a solution that treats the product name as highly relevant treats its specifications as highly relevant too. Specifications also share the same scoring budget as the words of the phrase, which stops a long phrase full of specifications from crowding out everything else.
You can see the result in two places:
- The playground shows what a specification contributed to a product's score, as a Technical Specification row in the score breakdown.
- The pipeline analysis shows the detected specification as the first step, before the text analysis steps.
What is not covered yet
The first version is deliberately narrow:
- Only the product name is read. Other text, such as the SKU name, the brand, or text attributes, is not.
- The set of recognized specifications is small. It holds the most common shapes for power, voltage, volume, weight, length, diameter, threads, dimensions, ingress protection, sockets, oil grades, and counts, and it grows over time.
- A unit spelled out in words, such as
10 meterinstead of10 m, is not recognized. - Count words, such as
pcsorstk, are recognized in Danish and English only. - Decimal and thousands separators that differ between languages are not unified, so
30.5and30,5are not yet treated as the same value. - A product gains its specifications the next time it is indexed. Products that are not updated keep being found the way they are today.