FootprintRAG: Visual Analytics for Evidence Context Refinement in RAG-based Scientific Literature Exploration

arXiv cs.GR details FootprintRAG, an LLM-agent-powered visual analytics system for refining evidence context in RAG-based scientific literature exploration. The core change is simple but important: instead of treating retrieval, reranking, filtering, and synthesis as a black box, the system exposes the evidence set as something users can inspect and revise before generation.

FootprintRAG parses papers into text and figure evidence units, expands a query into parallel variants, then retrieves and assesses evidence across iterative rounds. It also surfaces ERS-ranked supplementary candidates from the broader corpus, giving users a way to compare retrieval directions and recover material that may have been dropped earlier in the pipeline.

The practical angle for game developers is strongest for teams building internal research assistants, design knowledge bases, or production-facing LLM tools. If your workflow depends on trustworthy summaries, provenance-aware review and evidence tracing can reduce hallucination risk and make it easier to defend why a system surfaced a given answer.

The authors evaluated the system with two case studies, a user study, and a workflow-level comparison against representative RAG systems. They say it helps users revise candidate evidence, trace summaries back to supporting units, and spot overlooked material more reliably than conventional hidden-retrieval setups.