AI Research Partner: An AI-Powered Web Platform for Research Paper Analysis, Summarization, and Ideation

Abstract

The rapid growth of published research literature has made manual, unaided reading a bottleneck for students and early-stage researchers, who must extract structured understanding from unstructured PDF documents while operating at varying levels of comprehension. This paper presents the AI Research Partner, a full-stack MERN (MongoDB, Express.js, React, Node.js) web platform that unifies the research-reading workflow — comprehension, synthesis, and ideation — into a single authenticated system. The platform ingests a PDF, extracts its text, and uses the Google Gemini large language model to generate multi-level (basic, medium, technical) section summaries, an interactive D3.js concept knowledge graph, novelty-rated research ideas, citation recommendations, auto-generated quizzes, and abstract/slide drafts, while a Socket.io-based real-time layer enables collaborative annotation among multiple users. The system was implemented end-to-end, evaluated through functional testing across eight modules, and benchmarked for AI feature response latency and concurrent-user scalability. Results indicate pass rates above 87% across all modules, typical AI response times of 3-17 seconds depending on feature complexity, and stable real-time note-broadcast latency under load, demonstrating that a single, prompt-engineered platform can reasonably reproduce the core stages of expert research reading within one coherent, collaborative interface.

Introduction

The text presents AI Research Partner, an integrated web platform designed to help students and early-stage researchers read, understand, analyze, and extend academic research papers. The motivation is that academic reading is more complex than ordinary reading because it requires comprehension, synthesis, and ideation, often across multiple disconnected tools.

Existing tools such as PDF readers, reference managers, and generic AI chatbots address only parts of this workflow. They generally do not provide persistent, paper-specific knowledge, adaptive levels of summarization, concept visualization, research-idea generation, self-assessment, and collaborative annotation in one system.

Main Problem

The system addresses the difficulty researchers face when they need different levels of understanding from the same paper. A reader may initially need only a simple overview but later require detailed understanding of methods, equations, results, and limitations. Static abstracts and conventional PDF readers cannot adapt to these changing requirements.

Proposed Solution

AI Research Partner combines several research-support functions in a single authenticated MERN-stack application powered by Google Gemini. Its major features include:

• PDF upload and text extraction
• Multi-level summarization
• Knowledge-graph visualization
• Research-idea generation
• Citation recommendation
• Quiz generation and self-assessment
• Abstract and presentation-slide generation
• Paper-specific question answering
• Real-time collaborative annotation

The system stores generated AI artifacts so that they do not need to be regenerated every time a user revisits a paper, reducing unnecessary processing, cost, and latency.

Related Work

The research builds on several areas of previous work:

• Transformers and large language models provide the underlying language-model technology.
• SciBERT and Longformer demonstrate approaches for scientific and long-document processing.
• PEGASUS and discourse-aware summarization motivate section-specific summarization.
• SPECTER and citation-recommendation research inform the citation feature.
• GPT-3 and few-shot prompting support the decision to use prompting rather than training a separate model for each feature.
• Research on LLM-generated ideas motivates evaluating both novelty and feasibility.
• TextRank and graph-based methods provide conceptual foundations for representing relationships between concepts.
• Jupiter and CoAuthor influence the real-time collaboration design.

System Architecture

The platform uses a three-tier architecture:

1. Presentation layer: React 18, Vite, Zustand, React Router, Tailwind CSS, Framer Motion, and D3.js.
2. Application layer: Node.js and Express.js with REST APIs and Socket.io for real-time communication.
3. Data layer: MongoDB with Mongoose.

External services include Google Gemini for generative AI and Cloudinary for PDF storage.

The overall workflow is:

User registration → PDF upload → text extraction → AI processing → persistent storage → summaries/graphs/ideas/citations/quiz → Q&A and collaboration.

Database and API Design

The database centers around a Paper document. User information, papers, summaries, ideas, citations, knowledge graphs, notes, quizzes, and chat history are organized so that paper-specific information remains persistent.

The REST API provides endpoints for authentication, paper management, AI features, Q&A, notes, ideas, and research-paper generation. JWT authentication protects paper-related operations, while ownership and collaborator permissions control access.

Core AI Features

1. Multi-level summarization:
Each paper is divided into sections such as Abstract, Introduction, Methods, Results, and Conclusion. Each section receives basic, medium, and technical summaries, allowing users to choose their desired depth of understanding.

2. Knowledge graph:
Gemini extracts concepts, methods, results, applications, and limitations and represents their relationships as a graph. D3.js then provides interactive visualization.

3. Research Idea Generator:
The system produces 3–5 research ideas with methodology, expected outcomes, resources, and ratings for novelty, feasibility, and AI relevance.

4. Citation recommendation:
Relevant literature is identified using the paper's keywords and abstract, with recommendations organized according to relevance.

5. Quiz generation:
Gemini creates multiple-choice questions from the paper, while user attempts and scores are stored for later review.

6. Abstract and slide generation:
The system can transform the original paper or a generated research idea into a customizable abstract and structured presentation content.

Real-Time Collaboration

The Collaborate module uses Socket.io rooms. Users working on the same paper join a paper-specific room, allowing notes, typing indicators, and other events to be broadcast only to collaborators working on that paper.

The system intentionally uses a relatively simple last-writer-wins approach rather than complex operational-transformation mechanisms. This is considered appropriate because the goal is lightweight research annotation rather than simultaneous rich-text document editing.

Conclusion

This paper presented the AI Research Partner, a MERN-stack platform that unifies PDF ingestion, multi-level Gemini-powered summarization, an interactive knowledge graph, rated idea generation, citation recommendation, quiz-based assessment, and real-time collaboration inside one authenticated system. Functional testing achieved pass rates above 87% across all modules, response-time analysis showed sub-linear latency growth with document length, and load testing confirmed acceptable real-time performance at up to fifty concurrent collaborators. The prompt-generate-parse-persist pipeline kept seven distinct AI capabilities maintainable within a single service module without any task-specific model training, demonstrating that a general-purpose, prompt-engineered large language model can reasonably substitute for a collection of separately trained, task-specific models in this domain. Future work includes adopting schema-constrained generation to reduce JSON-parsing failures observed for the Knowledge Graph and Citation features, grounding citation recommendations in a real academic search API such as Semantic Scholar or CrossRef rather than the language model\'s own suggestions alone, replacing the last-writer-wins note broadcast with a conflict-free replicated data type for richer concurrent editing, moving the real-time room state to a shared store such as Redis to support horizontal scaling of the collaboration layer across multiple server instances, and extending PDF ingestion with optical character recognition to support scanned or image-based papers.

References

[1] A. Vaswani et al., \"Attention Is All You Need,\" in Proc. NeurIPS, 2017. [2] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, \"BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,\" in Proc. NAACL-HLT, 2019. [3] I. Beltagy, K. Lo, and A. Cohan, \"SciBERT: A Pretrained Language Model for Scientific Text,\" in Proc. EMNLP-IJCNLP, 2019. [4] I. Beltagy, M. E. Peters, and A. Cohan, \"Longformer: The Long-Document Transformer,\" arXiv:2004.05150, 2020. [5] J. Zhang, Y. Zhao, M. Saleh, and P. J. Liu, \"PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization,\" in Proc. ICML, 2020. [6] A. Cohan et al., \"A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents,\" in Proc. NAACL-HLT, 2018. [7] A. Cohan, S. Feldman, I. Beltagy, D. Downey, and D. S. Weld, \"SPECTER: Document-level Representation Learning using Citation-informed Transformers,\" in Proc. ACL, 2020. [8] C. Bhagavatula, S. Feldman, R. Power, and W. Ammar, \"Content-Based Citation Recommendation,\" in Proc. NAACL-HLT, 2018. [9] J. Beel, B. Gipp, S. Langer, and C. Breitinger, \"Research-Paper Recommender Systems: A Literature Survey,\" Int. J. on Digital Libraries, vol. 17, no. 4, pp. 305-338, 2016. [10] T. B. Brown et al., \"Language Models are Few-Shot Learners,\" in Proc. NeurIPS, 2020. [11] C. Si, D. Yang, and T. Hashimoto, \"Can LLMs Generate Novel Research Ideas? A Large-Scale Human Study with 100+ NLP Researchers,\" arXiv:2409.04109, 2024. [12] Q. Wang, D. Downey, H. Ji, and T. Hope, \"SciMON: Scientific Inspiration Machines Optimized for Novelty,\" in Proc. ACL, 2024. [13] M. Heilman and N. A. Smith, \"Good Question! Statistical Ranking for Question Generation,\" in Proc. NAACL-HLT, 2010. [14] X. Du, J. Shao, and C. Cardie, \"Learning to Ask: Neural Question Generation for Reading Comprehension,\" in Proc. ACL, 2017. [15] R. Mihalcea and P. Tarau, \"TextRank: Bringing Order into Texts,\" in Proc. EMNLP, 2004. [16] D. A. Nichols, P. Curtis, M. Dixon, and J. Lamping, \"High-Latency, Low-Bandwidth Windowing in the Jupiter Collaboration System,\" in Proc. ACM UIST, 1995. [17] M. Lee, P. Hoffman, and M. O. Riedl, \"CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities,\" in Proc. CHI, 2022. [18] Google, \"Gemini API Documentation,\" ai.google.dev, accessed 2025. [19] MongoDB Inc., \"Mongoose ODM Documentation,\" mongoosejs.com, accessed 2025. [20] Socket.io, \"Socket.io Documentation,\" socket.io/docs, accessed 2025.

Copyright

Copyright © 2026 Chintha Kameswara Lokesh, Dr. A. S. N. Chakravarthy, Priya Darshini Cholla. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.