Removing Words to Improve Results

Why remove words?

Removing words is the progressive loosening of query constraints to include more results when none are initially found.

For example, imagine an online shop that sold a limited inventory of iPhones in only 16 GB and 32 GB varieties. Users searching for “iphone 5 64gb” would see no results. This isn’t ideal behavior; it would be far better to show users some iPhone 5 results instead of a blank page. That’s where query expansion comes in.

Remove words if no results

The removeWordsIfNoResults parameter helps you make an initial query less and less specific until results are found. The right value to choose for a given use case may depend on the language searched as well as usage patterns.

You can choose one of these four behaviors: none (default), lastWords, firstWords, allOptional.

none

This is the engine’s default behavior. The engine doesn’t do any additional processing when a query has no results.

lastWords

This value treats a query’s last word as optional, and if there are still no results, it repeats the operation until either there are results, or the beginning of the query string has been reached.

For example, imagine a search for sparkly blue iPhone cases:

• The first query is sparkly blue iPhone cases.
• The second query is sparkly blue iPhone.
• The third query is sparkly blue.
• The fourth query is sparkly.

firstWords

This value treats a query’s first word as optional, and repeats the operation until either there are results, or the end of the query string has been reached.

• The first query is sparkly blue iPhone cases.
• The second query is blue iPhone cases.
• The third query is iPhone cases.
• The fourth query is cases.

It’s important to consider typical search patterns when deciding between firstWords and lastWords. For example, firstWords would be more suitable than lastWords in queries similar to the preceding example. However, this isn’t always the case. For example, take a look at the query “iphone 5 32gb.”

Expanding “iphone 5 32gb” with firstWords

1
"iphone 5 32gb" ➡️ "5 32gb" ➡️ "32gb"

Here, the most relevant part of the query is actually at the front; discarding those words makes the query irrelevant. Compare to the use of lastWords:

Expanding “iphone 5 32gb” with lastWords

1
"iphone 5 32gb" ➡️ "iphone 5" ➡️ "iphone"

This is much better—stripping away detailed descriptors expands the result set without making the query irrelevant.

allOptional

If there are no results for the initial query, allOptional specifies a second search in which all words are treated as optional. This is essentially changing the implicit AND operator between words to OR:

1
"blue AND iPhone AND cases" ➡️ "blue OR iPhone OR cases"

This last option works exactly like optionalWords, which is sent at query time. See Creating a list of optional words for more information.

Removing words with non-alphanumeric characters

When your query contains non-alphanumeric characters, you may notice some unexpected behavior when changing removeWordsIfNoResults from the default setting of none.

Some non-alphanumeric characters, or “separators”, trigger special behavior in the engine. Specifically, they trigger the concatenation of the surrounding alphanumeric characters. For example, the query t-shirt, matches on tshirt as well as t shirt. Additionally, the search looks for the parts in order. Looking for particular parts in a specific order is called a sequence expression. This means the query t-shirt matches t-shirt, t shirt, or tshirt, but not shirt t.

Therefore, if you’ve enabled removeWordsIfNoResults, and a user searches for a term like XYZ-b5, you may expect the query to match records containing only XYZ, if there are no results for XYZ-b5. This isn’t the case; because of concatenation and the subsequent sequence expression, the query matches XYZ b5 or XYZb5, not only XYZ.

Please refer to the guide on searching in hyphenated attributes SKUs, ISBNs, phone, and serial numbers for best practices when performing this kind of search. The documentation also has in-depth guides on tokenization, splitting, and concatenation, if you’re interested in learning more on these topics.