With Elasticsearch and Solr, you can easily customize analysis and scoring. There are several scoring algorithms built into them [1] such as BM25 (the default now) which is considered the state of the art for keyword relevance. For analysis, you can remove stopwords, stem, apply synonyms, etc [2]. Elasticsearch is specifically designed to scale across multiple machines which is necessary for TB datasets. There are also things like "more like this" queries and context-aware spell checking. Some of that you can do with PG, but not all of it. If PG can do it, it is usually harder to set up.
Rich library of tokenizers and analyzers. A test proof analyzer model and pipeline. For full text search, different score modes are supported that is beyond the trivial case mentitioned tf-idf model, how are going to do field centric ranking in postgres?
As far as I am concerned, it is far superior to the goodies that mentioned in this article.
