GAP: A Browser-Only Toolkit for Evaluating Multimodal RAG

We present GAP, an open-source, zero-install, browser-only toolkit that automatically evaluates Multimodal Retrieval-Augmented Generation (MRAG) systems on real engineering formula reasoning. Users simply upload structured JSON predictions produced by any MRAG pipeline; GAP instantly computes three orthogonal metrics against 200 expert-crafted questions derived from 10 authoritative technical documents (standards, datasheets, and manuals): (1) Ground Formula: symbolic correctness of extracted LaTeX formulas, (2) Answer Accuracy: numerical or algebraic correctness of the final solution, (3) Provenance Accuracy: faithfulness of cited document and page. Evaluation of eight frontier multimodal MLLMs reveals a universal pathology: provenance accuracy exceeds 98%, yet formula grounding and answer accuracy collapse below 7%. Current MRAG systems reliably locate the correct page but almost always hallucinate or mangle the underlying mathematics. Running entirely client-side, GAP provides the first standardized, reproducible, and instantly accessible way to quantify symbolic reliability in quantitative engineering tasks. The code and online demo of the GAP MRAG evaluation toolkit can be found here.