Integrating a scalable exemplar retrieval corpus (RAG) for diagrams

[lingyue404]

One idea on the design space: have you considered plugging in a scalable exemplar retrieval corpus for methodology diagrams (i.e., RAG with retrieved diagram examples + their textual context), instead of relying only on prompt-based planning/refinement?

We released DiagramBank (≈89k schematic diagrams with caption/abstract/in-text context) and a working RAG implementation for indexing and top-k retrieval. Our hypothesis is that retrieved exemplars could help:

• stabilize layout/style choices,
• reduce hallucinated components,
• improve alignment between method text and final diagram.

If this is aligned with your roadmap, we’d be happy to share an adapter and run a small integration/ablation.

Links:
DiagramBank repo: https://github.com/csml-rpi/DiagramBank
Preprint: https://www.researchsquare.com/article/rs-8917857/v1