SPARK – Smart Plug-and-Play AI Framework for RAG & Knowledge

Abstract

The rapid evolution of Generative AI has demonstrated strong capabilities in natural language understanding; however, its capabilities and applications in specialized domains remain limited due to issues such as hallucinations, a lack of contextual grounding, compliance concerns, and privacy concerns. This paper presents framework SPARK, which is Smart, Plug and Play AI framework based on RAG and Knowledge. RAG systems are designed to enable domain-specific assistants that are privacy-aware, deploymentflexible, and require no model fine-tuning. The proposed system supports multiple data modalities and inference backends, e.g., AWS Bedrock, Google Vertex, Azure AI Foundry, or GPUenabled, while offering a fallback mechanism that integrates private knowledge bases with controlled Internet search. The framework also augments the system by engineering the prompts on-the-fly i.e. Prompt Engineering. As a Proof of concept, we demonstrate the use case of a student and professorcentric AI assistant deployed in a university setting, grounded in instructor approved course material and designed to provide custom course guidance. We detail the system design, deployment pipeline, and multi-layered query flow, concluding with a discussion of border applications in education, enterprise, and healthcare.

Ähnliche Arbeiten