Retrieval-Augmented Large Language Models as Components of Cognitive Computing architecture for Regulatory Knowledge Management

This study investigates the transformation of large language models (LLMs) from standalone generative tools into reliable components of cognitive computing infrastructure. By implementing a Retrieval-Augmented Generation (RAG) architecture, the author explores how locally hosted models can process complex regulatory information without relying on external cloud services. The research utilizes two Polish-language models, Bielik and PLLuM, deployed on consumer-class hardware (Apple M4) using Ollama and LM Studio. The system architecture incorporates a local vector database (ChromaDB) and prompt orchestration (LangChain) to ensure that generated outputs are grounded in verified legal documents.

The integration of RAG with local LLMs demonstrates a marked improvement in the reliability of legal text analysis. By providing the model with controlled access to external knowledge repositories, the system significantly reduces the risk of hallucinations—a critical requirement for regulatory compliance. The study confirms that this approach allows for dynamic updating of regulatory information without the need for computationally expensive model retraining. Furthermore, the use of local infrastructure ensures data sovereignty, as sensitive legal documents and internal organizational data never leave the secure on-premises environment.

For organizations operating in highly regulated environments, the ability to perform accurate, auditable, and private legal analysis is paramount. This research provides a practical blueprint for deploying advanced AI capabilities that balance the power of modern LLMs with the strict security and interpretability requirements of legal and administrative sectors. By demonstrating that high-performance cognitive computing can be achieved on standard hardware, the study lowers the barrier to entry for institutions seeking to modernize their regulatory knowledge management processes.