Metadata-Enhanced RAG with SLMs for Industrial Tabular QA

Key Points

• To systematically assess the performance of search-augmented generation (RAG) in industrial data contexts where schemata and metadata are crucial.
• Utilized the NREL FIED dataset to construct 2,510 QA benchmarks.
• Compared performance of small language models (SLMs) like Phi-3-Mini and Gemma-2B-IT with large language models (LLMs) such as GPT-4o and Claude-3.5-Sonnet.
• Analyzed metadata integration effects on response reliability and contextual accuracy.
• Meta-RAG significantly improved answer reliability and contextual precision across models.
• Performance enhancements were observed even in smaller language models (SLMs).
• The study highlights the practical applicability of SLM-based RAG pipelines in industrial settings.

Abstract