You need something in the data stream to tell you where the boundaries are.
For example, HTML separates the header from the body with a blank line. In the first instance, you just accumulate data until that is seen (and save whatever comes after it for later), then parse the header.
Another way is to prefix each message with a length, so you can read say 1 or 2 bytes initially, then know how many more to expect for the next 'frame' of data.
