Frugal Knowledge Graph Construction with Local LLMs: A Zero-Shot Pipeline, Self-Consistency and Wisdom of Artificial Crowds

The project is a very early-stage research artifact (1 day old, 0 stars) accompanying an Arxiv paper. While the 'Frugal' angle (running on consumer hardware) is pragmatically useful, the core methodology relies on known techniques like self-consistency and model ensembling applied to existing benchmarks (DocRED, HotpotQA). Defensibility is nearly non-existent as the code serves primarily to validate the paper's findings rather than provide a persistent tool or service. From a competitive standpoint, this project faces extreme pressure from 'GraphRAG' initiatives by major players like Microsoft and various startups (e.g., WhyHow.ai, Unstructured) that are productizing automated KG construction. Frontier labs (OpenAI/Google) are also making this obsolete by improving native 'JSON mode' and long-context reasoning, which reduces the need for the complex multi-model 'Wisdom of Crowds' orchestration proposed here. The project's primary value is as a benchmark for how well small models (e.g., Llama 3 8B, Mistral) can perform relative to GPT-4 in structured extraction tasks.