PDF-SmartAI - An AI Study Companion for Uploaded Textbooks: A Book-Grounded Question-Answering System Using In-Browser PDF Extraction and Google Gemini

Abstract

Students increasingly turn to general-purpose AI chatbots for help with their studies, yet these chatbots answer from what they absorbed during training rather than from the textbook a student is examined on, and they can state wrong facts with confidence. This paper presents PDF-SmartAI, a zero-cost web application in which a student uploads up to two PDF textbooks and asks questions that an assistant, AI Buddy, answers only from that book, citing the page numbers it used and saying so when the book does not cover a question. Text is extracted entirely inside the browser by PDFium compiled to WebAssembly, split into page-aware chunks, stored per user in Cloud Firestore under owner-only security rules, and sent to Google Gemini through Firebase AI Logic on free tiers — the whole book when it is short (up to 100,000 characters), otherwise the eight chunks chosen by a plain-Dart keyword retriever. The study is secondary research combined with the evaluation of a built artefact; no participants were recruited. A public global survey of higher-education students (analytic n = 12,909) shows a moderate positive association between how much students use ChatGPT and how useful they find it for learning (Spearman ρ = 0.376, p < .001), while a substantial share of students (44.3%) agree that ChatGPT might mislead with inaccurate information (Wilcoxon z = 21.286, p < .001). On 10,570 SQuAD v1.1 questions over 13 synthetic books, the app’s own retrieval code placed the answer-bearing passage among the eight excerpts sent to the AI for 94.9% of questions, far above a random baseline (4.9%) and within 0.05 of BM25 (95.7%), although statistically slightly below it. All three hypotheses are supported. AI answer quality was checked only through live smoke tests and remains to be benchmarked.