Hi Amal, if its a feed forward neural network that u intend to use, then u can only specify the values of inputs to the neurons in the input layer and output/response values to the neurons of the output layer. However one cannot specify values to the neurons of the hidden layer/s. The weighted inputs from input layer and the bias are summed up at each neuron of the hidden neuron. Then the computed output of the hidden nerons and the bias are weighted and summed up at each neuron of the output layer during the training stage. 

I think u can use vector space model algorithm.

Go through this paper.

http://in.mathworks.com/help/images/ref/activecontour.html

 I join Vijay, you can't specify the values of hidden layers neurons. However, you can use DSSM or CLSM neural network architectures to learn  a latent representation of document and queries then you can add a cosine layer on the top of the NN in order to improve the similarity between relevant documents.

http://research.microsoft.com/en-us/um/people/jfgao/paper/2013/cikm2013_DSSM_fullversion.pdf

https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/www2014_cdssm_p07.pdf