◇ Could this help me?
Yes. The history is accurate and so is the core claim that model-driven extraction removed the hardest barrier to building a knowledge graph, which is a genuinely useful thing to understand before choosing a grounding strategy. One correction to carry: a graph extracted by a model inherits that model's errors, so extraction quality needs its own validation pass rather than being trusted because a graph looks authoritative. This is a general practice.
I want to evaluate graph-based grounding against plain retrieval for a body of documents I will describe. Explain what entities and relationships you would extract, what the output structure would look like, and how you would validate that the extracted graph is actually correct rather than plausible. Then tell me honestly for my case whether the added complexity beats straightforward retrieval, and what would have to be true for the graph to be worth it.