OWLLESS: A Semi-Automated LLM Framework for Extracting Methodological Design and Network Data from Social Network Studies

Social network research heavily relies on name generators, yet the specific impact of methodological choices on collected data remains understudied also due to the high labor cost of manual meta-analysis. This study introduces, validates, and demonstrates the use of an open-weights, local Large Language Model (LLM) framework, called OWLLESS, that is designed to semi-automate the extraction of methodological characteristics and network statistics from social network literature. Prioritizing transparency and data security, the framework operates entirely locally, processing PDF documents through a modular visual extraction and reasoning architecture using open-weights models (Pixtral and Llama 4 Scout). We validated the framework against a manually coded `gold standard' dataset of 61 publications on children's friendship networks. The framework demonstrated competitive performance across 10 variables sharing a common context, achieving an F1 score of 0.74 for numerical extraction and an overall accuracy of 81%. Furthermore, a pilot application examining reciprocity demonstrated the framework's utility by showing initial evidence of the influence of network boundary specifications on observed reciprocity. By streamlining data extraction while adhering to open science principles, this tool provides researchers with a scalable foundation to generate comprehensive meta-analytical insights, facilitating a deeper understanding of how study design shapes social networks.