Search engines employ various neural methods to match the users’ queries with relevant documents. An impediment to this is that natural language is rich and context-dependent, leading to several ambiguities in query formulation, making disambiguation a considerable challenge. Search engines must be robust to these ambiguities to encode the query text into a latent representation accurately. Converting tokens into approximate representations risks compressing the information and reducing its richness. While contextual embeddings have improved the quality of these representations, they have yet to eliminate all information loss and ambiguity. This work aims to investigate the robustness of search engines to ambiguities of the query formulation process, both when the query is formulated from an abstract information need to a text and when encoding that text into a latent representation. The proposed framework mitigates four aspects of linguistic ambiguities: surface form, grammatical, language-based, and script-based. The proposed framework optimises both the text and latent representations of the queries, allowing users to search for their desired information by making it the search engine’s responsibility to serve the user’s needs, rather than burdening the user with the need to formulate the right query that the system implicitly requires. This will enable greater accessibility in information search, allowing a variety of users to search for information.