Pollution Remediation Technology Retrieval Platform: source code and evaluation evidence
Abstract
Software and evaluation evidence accompanying the manuscript "A traceable knowledge base and hybrid retrieval system for contaminated site remediation technology recommendation". Contents pollution-remediation-system-code-d0427db.zip — full source code snapshot (Python 3.10 FastAPI backend, Vue 3 frontend, PostgreSQL 15 + PostGIS + pgvector deployment, Docker configuration; commit d0427db of the companion repository). evaluation_evidence_v1.zip — anonymised evaluation evidence: pre-specified analysis plan, annotation guidelines and anonymised annotation files, per-query result files underlying every table, frozen-corpus manifests with restoration instructions, classification review sheet, inter-annotator agreement analyses, and latency/graph-instrumentation measurements. Reproduction procedure: see evaluation_release/README.md inside the archive. deposit_README.md — deposit overview. System summary: a knowledge-based system that organises contaminated site remediation technologies (111 national R&D project reports, 466 traceable engineering cases, 3,223 publications) and supports hybrid retrieval (semantic vector search, synonym-expanded BM25, knowledge-graph provenance) fused through parameter-free reciprocal rank fusion, with traceable technology recommendation. Zero-shot LLM classification: 87.1% accuracy (macro-F1 0.89). Retrieval evaluated at two corpus scales (1,107 vs 3,830 chunks) on 50 queries with TREC-style pooled relevance judgements (κ=0.93); independently authored 46-query replication; ablation, significance testing, and negative results reported.
