URIs form the vocabulary of the Semantic Web; RDF is its grammar. And it is a very simple grammar. Information is communicated in RDF as a set of simple declarative sentences of the form subject - verb - object, where these three entities are all named by URIs. (The object can also be a literal such as the number 5.) This is pretty much the whole story; there are some details involving pronouns (called "blank nodes") that enrich the language a bit but they do not alter the essential story.
RDF is a bit like mud in the way it combines with itself and can be broken apart. You can cut mud into pieces, and you still have mud. You can select a portion of mud, and the portion is mud. RDF has the same properties. This is one thing that makes it ideal for data and knowledge integration.
Of the two syntaxes we prefer Turtle because it is concise, readable, and easy to understand, at least relative to RDF/XML. It is also consistent with the SPARQL query language. Unfortunately its standardization status is somewhat murky, and there are some tools that will accept RDF/XML but not Turtle. However, Turtle is sufficiently stable and well defined that we believe it can be used with little risk. If text must be converted from Turtle to RDF/XML this may be done using triplr or another similar tool.
RDF may be embedded in XHTML and XML using RDFa.