Page-Evidence RAG: Text-First Retrieval with Original-Page Evidence for Document Question Answering

Page-Evidence RAG is a standalone PDF/TXT research library and command-line tool for text-first retrieval with original-page evidence in multimodal answers. It combines MinerU Pipeline, structure-preserving text chunks, BGE-M3 dense/sparse retrieval with ZVec, original-page rendering, and backend-validated evidence-ID citations. It includes no visual embedding index, generated image descriptions or image-retrieval enhancement.

The archive contains source code, synthetic offline tests, a technical report and sanitized exploratory experiment materials. In one 25-question run over five publicly sourced PDFs, 13/20 answerable cases hit a target page in Top-5, 15/20 had correct target-page evidence and citation with correct requested fields, and five negative cases avoided invented parameters. These are AI-assisted assessments from a small purposive, previously used test set, not independently adjudicated or peer-reviewed results. No text-only comparative baseline or held-out effectiveness evaluation was performed.

The answer endpoint reported the label gemini-3.8-flash; its underlying vendor revision was not independently verified. Third-party PDFs, extracted full text, page images, model weights, credentials and private infrastructure are not distributed. Source URLs and hashes are provided. Author: 张良. AI assistance in implementation, documentation and evaluation is disclosed in the report.