Locayta Search® - Ranking Algorithm

Locayta's text relevance algorithm is used in both our ESP product and the Locayta Search Mobile iOS library. It is derived from the powerful BM25 ranking algorithm, which produces a relevance score for each result as shown:

∑ t f t ( k + 1 ) f t + k ( bL − b + 1 ) ⋅ log max ( N − n t , 0 ) + 0.5 n t + 0.5

\(t\): each term in the query
\(f_t\): within document frequency of term \(t\)
\(k, b\): custom parameters per search
\(L\): normalised document length (document length divided by average document length)
\(N\): number of documents in the index
\(n_t\): number of documents containing term \(t\)

In short, the relevance score, and therefore ranking, of a document is based roughly on how many of the user's query words exist in the document, how common those words are in the document itself and how frequently they occur in the entire index.

For example, if one document contains the user's search word just once, it will likely be ranked below another document that contains the word many times. If the user's query contains two words, and the first word occurs more rarely in the index, it is considered the more specific word and results with that word will be prioritised. This means that the search engine is very good at finding documents that closely match the user's query.

When indexing, you can also change the weightings of different fields. For example, if a title field had a higher weighting than the summary field, the search engine would favour results which contained the user's words in the title field over results which contained the words in just the summary.

There are also some other mechanisms at work which assist and augment this ranking algorithm, such as stemming, spell correction, and the default operator. You can find out more about these on our Search Engine and Mobile Search feature set pages.

For more information, Locayta's Technical Overview page provides details on how Locayta's search engine solves specific search problems. The Wikipedia entry for Full Text Search may also be of interest.