We've been running a study on RAL (Retrieval Augmented Localization). The pattern is structurally identical to RAG: at inference time, decompose the source paragraph into n-grams, embed them, cosine similarity search against a glossary vector index,…