An approach to information retrieval is contemplated for facilitating semantic search and response over a large domain of technical documents is disclosed. First, the grammar and morphology of the statements and instructions expressed in the technical documents is used to filter training data to extract the text that is most information-rich, that is the text that contains domain-specific jargon, in context. This training data is then vectorized and fed as input to an SBERT neural network model that learns an embedding of related words and terms in the text, i.e. the relationship between a given set of words contained in a user's query and the instructions from the technical documentation text most likely to assist in the user's operations. There are two parsing tasks. The first is to select a minimal sample of sentences from the document corpus that capture the domain-specific terminology (jargon). The result is set of sentences used to train BERT and SBERT. The second parsing task to create a set of action-trigger phrases from the document corpus. The trigger potentially matches a user query and the action is the related task.