Implementation of a custom NLP model on the main stream of data coming from several sources with the objective of deduplicate records based on embeddings similarity