The rapid rise of unstructured text data in various scientific, technical and corporate contexts is driving the necessity for auto-mated tools to distill usable knowledge from this raw text without extensive curatorial effort. In this paper we introduce a framework which uses generative AI techniques to convert unstructured documents into structured knowledge bases that can be queried using a pipeline of large language models combined with ontology-based extraction, entity canonicalization and retriev-al-augmented generation. Our architecture solves three recurring problems of knowledge discovery: extraction hallucinations, entity fragmentation and lack of alignment between queries and evidence in extracted knowledge. Our framework uses a two-step extraction approach that distinguishes between candidate generation and structural validation, with an identity resolution component that unifies various forms of surface-level entities into canonical entities. Experimentally, on a wide range of hetero-geneous corpora, we show that our framework greatly mitigates noise introduced by node generation from unstructured docu-ments, improves consistency of entity resolution and enables multi-hop querying over document-driven knowledge structures.