Matching instances of claims to claims in the graph

When the document-to-claims agent identifies claims, to get value out of the claim graph — or to contribute in the right place without duplicating claims — it needs to be able to find that same claim in the graph. Since claims can be stated in any number of ways using different words, this can only plausibly be done using vector search. Merely taking the first hit won't work, however; one needs an LLM agent iteratively using a vector-search tool.

I have not populated the claim graph to such a size that I can definitively say the current code works perfectly; however, I have built similar vector-search RAG tools for agents over a large corpus in my day job, and they work well enough that I am confident it will scale. It is especially likely to work if agents landing on adjacent claims can read those and traverse the claim graph.