How to train a W2V model in different contexts?

Question

Python

Good day everyone.
There was a need to create a visualization of the semantic similarity of words using w2v technology within several hundred documents, each of which includes various topics and filters.
And, accordingly, we visualize, for example, a word cloud only for a certain geographical region of people of a certain age.
The simplest option that I see is to form training samples for all possible combinations in order to teach the model to "speak" the language of each of the possible options. The only problem here is that the final number of options is from 230k to 937kk, which in principle is very sad.
I am not immersed in technologies like text processing on the issue of searching for semantic relationships and visualization .. so I will be grateful for any direction in which it is worth digging.
Now I use for visualization: https://projector.tensorflow.org/
I implement the w2v technology itself in gensim Python