GRU autoencoer: How to obtain a latent space composed of a single vector if there are multiple samples?

Question

I am using a bidirectional GRU autoencoder to obtain a single vector that represents in a compressed form the input time series, in order to cluster them. These time series have about 20000 time steps and 100 features each, so I am dividing them into multiple subsequences (samples) of 200 time steps.

I obtain that the latent space generated by the autoencoder is not a single fixed length encoded vector, but a sequence with as many vectors as the number of samples, that is to say, each sample is represented by a different encoded vector. How can I do to obtain a single vector that models all the samples?

I have tried with a stateful LSTM, but the resulting latent space is still a sequence of vectors, so I am not sure if the last vector of this sequence contains the information of all the samples or only the last one.