In the era of the Internet, the search for specific information in huge amounts of numerical data has become challenging. Some search engines and information retrieval methods have addressed this problem, yet only recommendation system engines provide the user personalization needed in most domains. In the field of aeronautics and defense particularly, information retrieval and recommendations are not that simple. The S1000D norm, used to standardize the drafting of technical documentation, is a complex XML-based norm with many rules and regulations. Until today, and to our knowledge, the only available search engine for S1000D documents is basic and relies on searching for the document that has the most occurrence of a query. This frustrating result leaves users with the same list of documents, having very little relevance to what they want.This research, in collaboration with Studec, a pioneer in technical documentation, focuses on enhancing and simplifying the retrieval of relevant S1000D documents to users. We address important limitations in the searches, including the importance of the relationship between user type and document type, in addition to the semantic meaning of the query and document. This is added while preserving the important rules behind the documentation, including the complex applicability, limiting access to data, and ensuring security. We also considered two different versions of the norm, having different architectures.The first step consists of preprocessing the data to simplify the form of the documents to be used in advanced models while keeping the meaning behind them, including the applicability filtering. We proposed a model that extracts the important information needed while preserving its applicability. We then converted the two versions of applicability into one form, filtering them using AND/OR Trees.The second part consists of retrieving and recommending relevant documents. The candidate generation phase consists of filtering the dataset by applicability and then retrieving documents that are either similar to a user's query or to his history. Documents are then reranked considering their type's importance to the user's job, and their importance to his previous searches. We used XLNet model to create text embeddings for semantic meaning for the first phase and created the deep neural network with an attention mechanism to rerank the extracted documents based on their relevance to the user’s job and history. Our final model is the first S1000D intelligent retrieving model that tackles not only semantic and fuzzy query searches but also weights the relevant documents based on the user’s profile and history.