During a recent client visit we encountered a common problem in search – over-application of ‘boosts’, which can be used to weight the influence of matches in one particular field. For example, you might sensibly use this to make results that match a query on their title field come higher in search results. However in this case we saw huge boost values used (numbers in the hundreds) which were probably swamping everything else – and it wasn’t at all clear where the values had come from, be it experimentation or simply wild guesses. As you might expect, the search engine wasn’t performing well.
A problem with both Solr, Elasticsearch and other search engines is that so many factors can affect the ordering of results – the underlying relevance algorithms, how source data is processed before it is indexed, how queries are parsed, boosts, sorting, fuzzy search, wildcards…it’s very easy to end up with a confusing picture and configuration files full of conflicting settings. Often these settings are left over from example files or previous configurations or experiments, without any real idea of why they were used. There are so many dials to adjust and switches to flick, many of which are unnecessary. The problem is compounded by embedding the search engine within another system (e.g. a content management platform or e-commerce engine) so it can be hard to see which control panel or file controls the configuration. Generally, this embedding has not been done by those with deep experience of search engines, so the defaults chosen are often wrong.
The balance of relevance versus recency is another setting which is often difficult to get right. At a news site we were asked to bias the order of results heavily in favour of recency (as the saying goes, yesterday’s newspaper is today’s chip wrapper) – the result being, as we had warned, that whatever the query today’s news would appear highest – even if it wasn’t relevant! Luckily by working with the client we managed to achieve a sensible balance before the site was launched.
Our approach is to strip back the configuration to a very basic one and to build on this, but only with good reason. Take out all the boosts and clever features and see how good the results are with the underlying algorithms (which have been developed based on decades of academic research – so don’t just break them with over-boosting). Create a process of test-based relevancy tuning where you can clearly relate a configuration setting to improving the result of a defined test. Be clear about which part of your system influences a setting and whose responsibility it is to change it, and record the changes in source control.
Boosts are a powerful tool – when used correctly – but you should start by turning them off, as they may well be doing more harm than good. Let us know if you’d like us to help tune your search!
Good caution, but for articles you should always boost the title with special care. Disabling frequency normalization for short fields like titles is a good idea, since otherwise you tend to penalize the existence of leading “the” and so on. Another thing that’s often poorly accounted for in default set ups is people’s expectation that an “exact match” to a title should always trump pretty much anything else.