Ontology-Guided Extraction Builds Knowledge Graphs from Documents

Large language models (LLMs) are adept at extracting entities and relationships from unstructured text, but they often produce inconsistent results, leading to fragmented vocabularies, duplicate entries, and conflated identities. This research introduces a robust, production-ready extraction layer designed to transform live streams of heterogeneous documents into a validated knowledge graph, meticulously aligned with a formal ontology. The system processes various document formats, including PDFs, spreadsheets, Office files, and images, routing them through specialized handlers. It employs a locally hosted Qwen3.5-9B model, fine-tuned on the target ontology, for a two-pass extraction of entities and relationships. A key innovation is its ontology-guided extraction mechanism: a relevant subset of the curated ontology is dynamically retrieved via embedding similarity and injected into the LLM prompt, drastically reducing catalog overhead compared to static domain slices. Following extraction, results undergo a five-stage refinement pipeline. This includes deterministic cleaning, merging across document chunks, a second pass for relationships, six distinct deduplication algorithms that operate without further model inference, and an embedding resolution subsystem with a conflict guard. Evaluation on intelligence corpora demonstrated a significant improvement in search recall from approximately 70% to 95% without false merges, while also rectifying seven classes of silent quality defects, from text truncation errors to systematic entity duplication.