Irfhanna Ameer Bahadur Ibrahim Kalifullah

• Recent progress in generative artificial intelligence, particularly in large language models (LLMs), has enabled significant advances in natural language processing across many application areas. In expert-oriented systems, however, these models continue to exhibit critical limitations, including hallucinated content, factual inaccuracies, and reduced robustness when applied to specialized or technical domains. Retrieval-Augmented Generation (RAG) has emerged as a promising approach to mitigate these issues by explicitly grounding model outputs in external knowledge sources. This thesis examines the design and empirical evaluation of a modular RAG pipeline intended for expert applications such as technical documentation assistance and process automation. The work specifically investigates whether combining multiple retrieval modalities ranging from dense vector-based retrieval to graph-based symbolic reasoning can improve the factual grounding, relevance, and explainability of generated responses. Particular attention is paid to how different retrieval strategies affect downstream generation quality in real-world technical settings. The architecture integrates interchangeable retrieval components, context augmentation strategies, and transformer-based language models, enabling systematic comparison across multiple retrieval configurations. An iterative design and evaluation methodology is adopted, encompassing architectural analysis, implementation of retrieval and generation modules, and performance assessment on domain-specific datasets. Evaluation is conducted using automated metrics, with a focus on faithfulness and contextual relevance as measured by the RAGAS framework. The results provide practical insights into when and how retrieval augmentation improves expert-oriented text generation. In particular, the findings highlight the importance of retrieval quality, effective filtering, and modular system design for building reliable, transparent, and deployable RAG-based AI expert systems.