Physics-Grounded Knowledge Base Construction for Altermagnets with Structure-Aware Document Reading and LLM Extraction: Altermagnet Talking Database

Altermagnetism has emerged as the third fundamental magnetic phase alongside ferromagnetism and antiferromagnetism, generating a rapidly growing literature that is increasingly difficult to curate by hand. Here we present a physics-grounded knowledge-base pipeline for altermagnetic materials that combines symmetry-based material seeding, layout-aware document reading, LLM-assisted extraction, auditing, and vector-database retrieval within a distributed inference stack. The workflow begins from crystallographically screened candidate materials, harvests relevant literature from arXiv, transforms papers into structure-preserving text representations, and finally builds validated records through either single-agent or crew-based extraction paths before serving grounded question answers over our knowledge base. Across benchmark tests at two model scales, our structured single and crew pipelines greatly outperform the bare LLM baseline in recovering paper-grounded values, with the smaller model remaining consistently faster. Our analysis shows that scientifically useful performance is driven primarily by structured retrieval and high-fidelity document reading, with agentic complexity acting as a secondary design choice rather than the central source of accuracy gains.