Microsoft Proposes PIKE-RAG, an Industrial RAG Framework

It notes that existing RAG (Retrieval-Augmented Generation) has focused heavily on simple text-based benchmarks, limiting its ability to handle industrial-site data that includes complex formats such as tables, charts, and diagrams.

PIKE-RAG is a framework focused on augmenting Specialized Knowledge and Rationale, and it proposes two core techniques.

• Knowledge Atomizing: Extracts and structures key information from diverse forms of knowledge sources
• Knowledge-Aware Task Decomposition: Breaks down tasks into steps for reasoning according to the difficulty and complexity of the question

The framework classifies question difficulty into four types and presents a roadmap that can address tasks in stages, ranging from simple rule-based queries to high-difficulty tasks requiring multi-hop reasoning.