Hybrid expert system and large language model (llm) architecture

Abstract

One implementation describes a computer-implemented system and method that integrates expert system techniques with large language models to generate controlled, domain-specific answers. It works by receiving user queries through an interface and analyzing them to identify the relevant domain and subdomain. The system then employs multiple domain-specific expert subsystems—using rule-based and case-based reasoning, structured knowledge bases, and ontologies—to retrieve information through both semantic vector and token-based searches. Candidate answers with confidence scores are generated and processed by specialized, fine-tuned LLMs that deliver natural language explanations and responses, while a weighting engine combines outputs and an output guardrail ensures compliance with regulatory, ethical, and domain-specific standards. A refinement module further adjusts answers by re-ranking, cross-referencing with current peer-reviewed material, and normalizing against historical or real-world data, with human-in-the-loop review available if needed.