What should a retriever actually retrieve?
Conventional RAG often treats fixed chunks as the retrieval unit. Multi-hop questions make that assumption fragile because the evidence needed to answer a question can be distributed across multiple pieces of context.
Compare reasoning units with conventional chunks.
I built a framework that evaluates reasoning-unit retrieval against standard chunk-based retrieval. The comparison focuses on retrieval quality, evidence coverage, context efficiency, and downstream answer quality.
Measure both coverage and cost.
A retriever can improve recall simply by returning more text, so evidence coverage alone is not enough. Context efficiency makes the tradeoff explicit by asking how much relevant evidence is recovered for the amount of context sent downstream.
Retrieval quality is a systems metric.
The project treats retrieval as part of an end-to-end reasoning system rather than a nearest-neighbor lookup. Better retrieval means giving the model the evidence it needs while controlling context growth and preserving answer quality.