If you are using the ParallelInputter shown above, you should indeed feed the input text + feature vector.
If you meant the vectors, yes they have 2 dimensions in depth.
yes sorry, the vector. I understood now 
the tf record will have just an array of numbers which are the full sentence representation and the word embedding will create that same array size concatenating each word if each word has the same vector depth as the calculated for the TFRecord vector.
I guess parallel input the ContactReducer must be axis= -1 right?
Yes, to concatenate in depth.
thank you so much.
one advice to improve the documentation would be.
- What is exactly the aligments are for?
- must be clarified the ark file its the source file and not a db.
- must be clarified each key matches to a full example
- I didn’t know the tfrecord its just an array of numbers like numpy which will be reshaped in the shape needed for the system
I hope this helps to improve the documentation 