The paper's dense seeding step encodes both query texts and concept texts (drawn from the concept store's name + textual description fields) with the all-MiniLM-L6-v2 sentence-embedding model unless otherwise noted. Embeddings are L2-normalized, and retrieval uses inner-product search, which on L2-normalized vectors is equivalent to cosine similarity. The output is a dense top-$m$ seed pool of concept candidates that is held fixed across strict-parity comparisons, so any difference between compared systems comes from the downstream graph policy rather than from a re-seeded candidate set.