Reflexion-Based Agentic Content Review: An LLM-as-a-Judge Framework for Lakehouse ECM

Abstract

Enterprise departments such as legal and marketing uses content management systems like SharePoint and Salesforce for collaboration on large volume of marketing campaign and legal contracts content. The content needs to be fact-checked, compliant, relevant, consistent and of high-quality, to ensure all this is a time-consuming, manual and error-prone process. This paper presents a databricks based automated content review framework that leverages Reflexion-agent based agentic architecture where LLMs (large language models) play the role of judges to evaluate, critique and refine the content. Enterprise content ranging from marketing campaign articles to legal contracts are ingested into databricks Lakehouse using scalable, robust and self-healing ingestion pipelines.  The proposed Reflexion agent is configured using a powerful LLM, detailed system prompts with clear rubrics to ensure clarity, compliance and brand adherence by generating critiques and self-reflections iteratively to finally achieve the improved outcome without human involvement. Experimental results indicate accurate and automated content review across marketing and legal content repositories. This work highlights the potential of Reflexion agent-based content review systems and establishes Databricks as a platform for intelligent content life cycle management.