In emerging scientific domains, local Small Language Models (SLMs) are gaining traction because they offer stronger privacy guarantees and more stable deployment pipelines than large models. In practice, scientific QA with SLMs faces constraints such as limited literature collections, fragmented evidence, short context windows, and modest reasoning capabilities. To address these issues we introduce the Evidence‑Grounded Typed Knowledge Graph (EGT‑KG) retrieval framework, which enhances information retrieval for local SLMs. We evaluate three QA pipelines: a vanilla Retrieval‑Augmented Generation (RAG) workflow and two EGT‑KG variants that use an automatically generated relation schema (AS) and an expert‑defined relation schema (ES). Experiments on a Biopolymer‑bound Soil Composite literature benchmark employ a six‑dimensional evaluation framework (S3CRF: Soundness, Correctness, Completeness, Conciseness, Relevance, Fluency). EGT‑KG consistently outperforms vanilla RAG, with the strongest gains observed for llama3:8b, where AS and ES variants achieve final scores of 70.37 (+14.67%) and 68.82 (+12.14%) respectively.
Review