The Python version of the recommendation API uses navigation vectors to compute cosine similarity between Wikidata items. In order to do so, a big matrix (over a GB in size) needs to sit in memory and be used for computation. We can do better and pre-compute consine similarities and serve them from a database. This helps us get rid of the second instance of the recommendation API service and use the first instance for serving both translation and related articles based recommendations. We can later use the pre-computed data in the JS version of the recommendation API once T193746 is resolved.