FPN-RAG: A stratified retrieval-augmented generation framework for evidence-traceable technology hotspot forecasting in optoelectronics

Technology hotspots in optoelectronics evolve through coupled changes in materials, device and waveguide structures, fabrication routes, performance metrics and applications. Forecasting them requires evidence linking reported optical principles and structures to recent technical developments and emerging signals because the task is shaped by documented, often incremental, development rather than sudden unrecorded breakthroughs. Most retrieval-augmented generation (RAG) systems retrieve or summarize published knowledge but rarely impose retrospective evidence cutoffs, distinguish foundational from cutting-edge evidence, or require forecasts to specify a technical route, key constraint and application target. Here, we propose feature pyramid network-inspired retrieval-augmented generation (FPN-RAG), a prediction-oriented RAG framework that adapts the feature pyramid principle to scientific literature retrieval. It organizes evidence across time scales and technical functions and redesigns retrieval for optoelectronic hotspot forecasting through stratified, signal-guided evidence construction. FPN-RAG integrates a topic-constrained Web of Science (WOS) offline corpus with Bohrium online supplementary evidence. Candidate pools are characterized by semantic signals, evidence velocity, historical persistence and maturity (SVH/M), and dynamic quotas. Adaptive reranking and core-technology retention construct the final context. In a cutoff-controlled retrospective benchmark with year-specific reference hotspots for hollow-core fiber, perovskite solar cells and optical neural networks, FPN-RAG achieved a hit rate of 93.3% at the 0.6 threshold (HR@0.6), a hit rate of 31.1% at the 0.7 threshold (HR@0.7), and a mean semantic similarity of 0.676, outperforming the best non-FPN baseline for each metric by 4.4%, 11.1% and 0.017, respectively. These results indicate that FPN-RAG provides a training-free retrieval approach that maintains traceable evidence selection while producing better-aligned, structurally explicit forecasts from multisource evidence.