Can routing queries to task-matched structures improve RAG reasoning?

Knowledge-intensive reasoning tasks require useful information that is badly scattered across documents. Standard RAG approaches retrieve text chunks and feed them to the model — a uniform structure regardless of task type. StructRAG argues this ignores a well-established cognitive science finding: humans use different structured knowledge representations for different task types, and performance improves when structure matches task demands.

The framework applies two cognitive theories directly:

• Cognitive load theory (Sweller 1988): humans summarize scattered information into structured knowledge to shorten reasoning paths and enable more accurate judgment
• Cognitive fit theory (Vessey 1991): different structure types suit different tasks — tables for statistical analysis, graphs for long-chain inference

StructRAG implements this through three modules: (1) a hybrid structure router selects the optimal structure type from five candidates — table for statistical tasks, graph for long-chain tasks, algorithm for planning tasks, catalogue for summarizing tasks, and chunk for simple single-hop tasks; (2) a scattered knowledge structurizer converts raw documents into the selected format; (3) a structured knowledge utilizer infers answers from the resulting structure.

The router is trained via DPO on synthetic preference data generated through a task-synthesis → solution-simulation → preference-judgment pipeline. This addresses the data scarcity problem: real-world training data for "which structure type works best for this query" barely exists, so the system creates it.

This is distinct from existing graph-vs-vector RAG work. Since When do graph databases outperform vector embeddings for retrieval?, the existing insight is "use graphs for relational queries." StructRAG's insight is broader: route to any of five task-appropriate structure types including tables, algorithms, and catalogues — graph is just one option. Since Can reasoning topologies be formally classified as graph types?, there's a structural parallel: just as reasoning can be routed to different topology types, retrieval can be routed to different knowledge structure types.

The cognitive science grounding gives this theoretical backing beyond engineering heuristics. It suggests the principle generalizes: any time AI systems can represent the same information in multiple structural formats, routing to the task-appropriate format should outperform any single universal format.

Source: Routers