The semantic layer in retrieval and knowledge graph design

Why vector search alone falls short

Embedding similarity finds passages that read like the question. It does not encode that a study has sites, that sites have investigators, and that investigators have prior enrollment performance. The question "which sites should we consider" requires traversing all three.

Multi-hop questions need structure, which is what a knowledge graph provides. The stronger enterprise implementations are hybrid: graph for relationships, vector for language, both governed.

What I built at Parexel

A knowledge-graph and LLM pipeline joining structured clinical metadata to unstructured protocol documents through a RAG architecture, so operational teams could ask questions in natural language and receive governed, auditable answers about a global study portfolio. It covered intelligent search, protocol summarization, and site and investigator insight, and reduced analyst research time by around 60%.

The determining factor was not the retrieval strategy. It was that clinical metadata had already been mastered and the semantic relationships were explicit.

Design rules I hold to

Agents query the semantic layer rather than the physical schema. Extraction, classification, and embedding are governed pipeline stages that inherit lineage and policy in the same way as tables. Retrieval accuracy is measured with an eval harness per use case and tracked over time, because retrieval quality degrades quietly as a corpus grows.