Hardware datasheets contain engineering specifications required for component selection, system integration, and compatibility verification. The documents are long, heterogeneous, and spread specifications across tables, figures, and prose, making manual inspection time‑consuming and error‑prone. Although large language models (LLMs) can automate extraction, a naïve one‑shot approach often omits relevant specs even when the whole document fits within the model’s context window.\ \ We present an LLM‑driven pipeline that converts hardware datasheets into a general‑purpose hierarchical hardware knowledge graph. The pipeline consists of three key stages:\
- Chunk‑based property extraction – the document is split into manageable chunks; each chunk is processed by an LLM to produce key‑value property pairs.\
- Hierarchical routing – extracted properties are semantically mapped to appropriate graph layers, building a multi‑level node structure.\
- Iterative refinement – the preliminary graph undergoes multiple rounds of LLM reasoning and validated instruction execution; a deterministic local executor updates the graph while preserving source provenance.\ This staged design reduces per‑query context size and keeps intermediate representations traceable. The resulting knowledge graph is semantically organized, source‑grounded, and ready for downstream retrieval and reasoning.\ \ We evaluate the framework on 20 hardware datasheets covering diverse component types and document lengths. Property faithfulness measures how faithfully extracted attributes reflect the original text, while inter‑graph completeness quantifies relative semantic coverage without exhaustive ground truth. The framework achieves 98.8% property faithfulness versus 76.1% for the one‑shot baseline, and fully covers 83.6% of the baseline’s faithful properties; conversely, the baseline only covers 10.5% of ours. These gains provide more reliable and comprehensive engineering information for downstream hardware design and verification, reducing the risk of missed constraints and incorrect decisions.\ \ Review