The simplest way to start out would be to import from CSV (assuming it's easy to dump users/books/tags from MongoDB to CSV), use that to train the model, then d

The simplest way to start out would be to import from CSV (assuming it's easy to dump users/books/tags from MongoDB to CSV), use that to train the model, then dump the recommendations from the model back to CSV (or using Python code, write them directly to MongoDB). That would give you an offline system with one-time recommendations. You could then run this in batch, on a schedule, to re-train as you get new tags/bookmarks and update the recommendations periodically.