NAVINDEX ties injection on legal docs with 56x smaller context and 25% lower cost

A study compares full-corpus injection against two structured retrieval modes for analyzing transactional legal documents: embedding retrieval (NAVEMBED) and LLM navigation over a compact index (NAVINDEX). On a 20-question benchmark, both retrieval methods tied with injection on semantic accuracy while significantly reducing token usage.

• NAVEMBED tied injection on 16 of 18 document-bound questions while attending to 17.3x fewer input tokens.
• NAVINDEX tied injection on all 18 questions with a 1.61x smaller total token footprint and approximately 56x smaller answering context.
• NAVINDEX achieved 25% lower dollar cost compared to the injection baseline.

The authors derive a closed-form caching-crossover rule, noting that cached injection is only cheaper when the corpus remains below roughly ten times the retrieval payload.