DecisionRAG Navigates RAG Without First-Stage Retrieval

DecisionRAG challenges the assumption that RAG must start with embeddings, but it is a new retrieval architecture—not a universal replacement for production indexes.

• –Shared-context candidate judgments improve relevance comparisons by up to 7.7 points over independent scoring.
• –The approach is fast because decision models choose branches without generating tokens.
• –Its biggest weakness is irreversible pruning: poor hierarchies or vague descriptions can eliminate the right evidence early.
• –“Retriever-free” mainly means no query-time dense retriever; the Natural Questions setup still uses embedding-based clustering to build its hierarchy.
• –Developers should view this as a promising hybrid or hierarchical RAG design for structured corpora, not a drop-in solution for every knowledge base.