PlanSightRAG: A Visual-First Multimodal RAG for Automating Question Answering and Compliance Checking for Civil Standard Plans
PlanSightRAG keeps the plan sheet visual, using layout and geometry as evidence instead of flattening drawings into OCR text.
The paper says the system retrieves and reasons over civil standard plan imagery with a Planner-Retriever-Auditor-Synthesizer pipeline and MaxSim heatmaps for traceability. Its benchmark covers 4,056 pairs from five state DOT standard-plan sets, totaling 1,898 pages. The reported zero-shot retrieval result is 91.47% Recall@5, with 91.40% on a held-out Michigan DOT corpus. In synthetic compliance tests, the Qwen2.5-VL-72B pipeline reaches 100% verdict accuracy only when given a pre-resolved rule threshold. ArXiv · AI/CL/LG's note
The paper says the system retrieves and reasons over civil standard plan imagery with a Planner-Retriever-Auditor-Synthesizer pipeline and MaxSim heatmaps for traceability. Its benchmark covers 4,056 pairs from five state DOT standard-plan sets, totaling 1,898 pages. The reported zero-shot retrieval result is 91.47% Recall@5, with 91.40% on a held-out Michigan DOT corpus. In synthetic compliance tests, the Qwen2.5-VL-72B pipeline reaches 100% verdict accuracy only when given a pre-resolved rule threshold. ArXiv · AI/CL/LG's note
score 4