Bibliographic Tracking with an LLM

Overview

A Python system that automates the generation of .docx bibliographic tracking documents from a JSON file of academic references (DOI, arXiv, GitHub, web, PDF). It pulls real metadata from CrossRef, OpenAlex and arXiv, and combines it with an analysis generated by a local LLM running on Ollama, with automatic detection of GPU (NVIDIA/CUDA, AMD/ROCm) or CPU.

The system has four operating modes: metadata only (no LLM), CPU analysis, GPU analysis, and initial setup. Each generated document includes general data, the IEEE reference, abstract, a content analysis (methodology, results, contributions, limitations) and how it relates to the researcher's project. It supports models such as Mistral, Phi-3 Mini and LLaMA 3, taking 30–90 seconds per document on GPU.