Basically BM25 is just a complicated math formula that assigns a score to a document for a particular query. This does work great but if you read about it, it really only takes you so far when it comes to a quality result.

You need to bring in other factors to the score, like in the page rank example of links to the document and links out from the document. This doesn't apply for everyone use case but its a very good digestible example and let's say number of clicks from other users in the past month or some form of rate of decay over time for usage. I am going to over simplify this next part. During the scoring process you can add these three features together to come up with your new score.

This is a great start but you aren't done yet because not all things are created equal you don't want to treat your features with the same weight. Depending on how much money and time you have (Also domain experts are important) you can come up with various approaches to finding the right weights to assign to each of these features before you add them up to a score for the document. Some use gold data, meaning for 100 queries for example (you will want to do a lot more) categorize the documents within the results you get back and use a support vector machine to have it come up with the weights you should use.

Another way to do this is looking from a demo like UI, make tweaks to the weights and iterate till its good enough. Doing the latter approach can be useful for a POC to internally prove the value of something, but you will have to be very careful if you introduce that to your external customers unless you very clearly set their expectations.

This question was first asked in How to Enhance Your Search Platforms webinar (8th of Feb). Please see here: How to Enhance Your Search Platforms