Answers are grounded in this paper and cite their sources.
This paper addresses the optimization of hybrid retrieval and context formation processes in generative artificial intelligence systems based on large language models (LLMs). Despite remarkable generative capabilities, current LLMs suffer from a static knowledge cutoff, a tendency to hallucinate, and a lack of access to external data sources. Retrieval-Augmented Generation (RAG) represents a paradigm shift, enabling the integration of external knowledge bases into the generation process; however, the effectiveness of RAG systems critically depends on retrieval quality and the construction of an optimal context window.This work proposes a suite of optimization methods comprising: hybrid lexical-semantic retrieval with adaptive weighting via Reciprocal Rank Fusion; a query reformulation algorithm with complex query decomposition and synonym expansion; two-stage context formation with cross-encoder re-ranking; and a mathematical model for optimal text chunk size selection. Experiments were conducted on a corpus of 5 research papers (552 chunks) using Precision@5, Recall@5, F1, MRR, and NDCG@5 metrics. Results demonstrate an improvement in F1 from 0.44 (Naive RAG) to 0.58 (Optimized RAG), a gain of 14 percentage points. Query reformulation provides the largest improvement for ambiguous queries (+51.1%) and multi-step queries (+29.3%). Scalability analysis confirms acceptable performance up to 65,000 chunks. Findings are consistent with leading international studies on Blended RAG, Self-RAG, and RQ-RAG, as well as Ukrainian research by Morozov O. V. and Donets V. V. on RAG with ChromaDB and LangChain
No takes yet. Share an insight, caveat, or question.
Oleksandr Denysiuk (2026) studied this question.
Synapse has enriched 4 related papers. Consider them for comparative context: