Can smaller models handle RAG filtering while larger models focus on synthesis?

HiFi-RAG separates the RAG pipeline into stages handled by models of different capability and cost: a fast cheap model (Gemini 2.5 Flash) does query reformulation, prunes irrelevant retrieved passages, and attaches citations, while the large expensive model (Gemini 2.5 Pro) is invoked only at the final generation step. This is a tiering pattern with a specific theoretical justification: filtering and citation are pattern-matching tasks where the smaller model is sufficient, while final synthesis is where the large model's reasoning matters most.

The design implies a richer view of "RAG" than a single retrieve-then-generate pass. Each intermediate decision — which query to expand, which passages to keep, which spans to cite — has its own optimal cost-quality point, and forcing the most capable model to do all of them wastes compute on tasks where it offers no marginal benefit. The hierarchy also produces a useful side effect: because filtering happens before generation, the large model receives a smaller higher-quality context, which improves its answer quality even setting cost aside.

The general principle is that RAG architectures should think in terms of decision granularity rather than uniform model deployment. The retrieval pipeline contains several distinct sub-decisions, and matching each to an appropriately-sized model produces both cheaper and better answers — a Pareto improvement that uniform RAG misses because it treats retrieval and generation as a single coupled act. This is the RAG-specific instance of Can small language models handle most agent tasks? — heterogeneous tiered architectures are the economic imperative whenever subtasks have different capability requirements.