- By:
- Kannan, Ramakrishnan ; Gautam, Ashish ; Patton, Robert M; Haas, Nicholas Q; Thomas, Todd M; Aimone, James; Potok, Thomas E
- Page Number:
- 1-9
- Book Title:
- 2026 Neuro Inspired Computational Elements (NICE)
- Publication Date:
- September 15, 2026
- Publisher Location:
- IEEE, New Jersey, United States of America
- Conference Name:
- Neuro Inspired Computational Elements (NICE)
- Conference Location:
- Atlanta, Georgia, United States of America
- Conference Sponsor:
- Institute of Electrical and Electronics Engineers
- View DOI Listing:
- https://doi.org/10.1109/NICE69539.2026.11567520
Abstract
We introduce a scalable system for constructing scientific Knowledge Graphs (KGs) from neuromorphic computing literature using open-source, lightweight Large Language Models (LLMs), and for leveraging these graphs as structured retrieval substrates in a GraphRAG framework. Our pipeline integrates extraction, merging, canonicalization, and clustering to transform unstructured text into a consistent, high-quality KG. We deploy models such as gpt-oss-120 b, 11ama3.3:70b, and gpt-oss-20b as LLM backends in a cost-effective distributed processing architecture. Applied to 2,396 documents, the system extracted over 921,545 triples with a 100% query completion rate. We show that the resulting KG supports multi-hop reasoning, explicit provenance, and robust retrieval that overcomes the limitations of standard vector-based RAG. Within a neuromorphic co-design context, the GraphRAG layer provides substantially deeper and more hardware-aware responses than naive LLMs, capturing relationships across algorithms, circuits, devices, and architectures. These results demonstrate that combining open LLMs with structured scientific knowledge yields a practical and explainable foundation for accelerating research in neuromorphic computing and other technically complex domains.
