What are the five structure types and which tasks does each one suit best?

This reads as a question about StructRAG's five knowledge structures, where the whole idea is that there's no single best way to organize retrieved knowledge — the right shape depends on what the question is asking the model to do. The five types are tables (for comparing entities across shared attributes), graphs (for relationships and multi-hop connections), algorithms (for procedural, step-ordered tasks), catalogues (for descriptive, list-like summarization), and plain chunks (for simple factual lookup where structure adds nothing). StructRAG's real claim is that a DPO-trained router picks the structure based on what the query demands, and that this beats feeding everything through one uniform retrieval pipeline. The grounding is borrowed from cognitive science: cognitive fit theory says reasoning is easier when the form of the information matches the form of the task Can routing queries to task-matched structures improve RAG reasoning?.

The deeper move worth noticing is that this is one instance of a recurring pattern in the corpus — match the *shape* of representation to the *shape* of the problem. You see the same logic in reasoning topologies, where chain-of-thought, tree-of-thought, and graph-of-thought map precisely onto path graphs, trees, and arbitrary directed graphs, and where a graph's ability to merge branches (in-degree greater than one) lets it do divide-and-conquer synthesis a tree simply cannot express Can reasoning topologies be formally classified as graph types?. Picking CoT vs. ToT vs. GoT is the same kind of decision as picking a table vs. a graph: the structure is not decoration, it determines what computation is even possible.

The routing instinct also shows up at the task-assignment level rather than the knowledge-representation level. Intelligent delegation frames the matching problem across eleven task axes — complexity, verifiability, reversibility, and so on — arguing that the foundational question is whether a task's outcome can even be evaluated before you decide who or what should handle it What makes delegation work beyond just splitting tasks?. StructRAG asks 'which structure fits this query?'; delegation asks 'which agent fits this task?' — both reject the one-size-fits-all default in favor of characterizing the work first.

Where it gets more interesting is that structure isn't only chosen, it can be *built* as reasoning unfolds. Knowledge Graph of Thoughts externalizes reasoning into iteratively constructed knowledge-graph triples and gets a 29% jump on hard GAIA tasks using a small model — the graph structure does work the model's weights don't have to Can structuring reasoning as knowledge graphs help smaller models solve complex tasks?. And on the spectrum between 'no structure' and 'fully formal,' the corpus repeatedly lands in the middle: partial symbolic augmentation beats both pure natural language and full formalization because it adds structure without losing meaning Why does partial formalization outperform full symbolic logic?, and semi-formal natural-language templates capture the discipline of formal verification — no skipped cases, no unsupported claims — without paying the cost of symbolic rigor Can structured templates replace formal verification for code reasoning?.

So the honest answer is that the cleanest 'five structure types, one task each' list comes specifically from StructRAG, and that's the note to open if you want the literal taxonomy. But the thing you might not have known you wanted: the same select-the-right-shape principle quietly organizes a much larger slice of the collection — from how reasoning chains are topologically wired, to how tasks get delegated, to how much formalism a problem actually needs. The five structures are a doorway into a broader bet the field is making: that matching form to task is doing more of the work than any single clever method.