When you want a model to reuse past decisions, you need a store and a retrieval method. The combination of retrieval with generation is now textbook. See Lewis et al. (2020): retrieval-augmented generation for knowledge-intensive NLP.
We store knowledge units in Postgres and use the pgvector extension for similarity search. For the graph side of the algorithm, the standard reference is Malkov and Yashunin: HNSW (approximate nearest neighbors). That hierarchical navigable small world work is what makes large nearest-neighbor queries feasible at scale in many deployments.
So what
Vector search is not magic recall. It is an index and a distance function. If the buyer does not know what is stored and how similarity is defined, the feature is a demo.