Chunk Size as an Experimental Variable in RAG Systems

Retrieval-Augmented Generation (RAG) systems are analyzed by experimenting with different chunk sizes to understand their impact on retrieval performance. The experiment uses a minimal RAG system, without a final generation step, to isolate the effects of chunking on a small knowledge base. Using very small chunks of 80 characters led to significant context loss and fragmented, unusable results. A medium chunk size of 220 characters produced more coherent results but failed to distinguish between closely related but distinct concepts, demonstrating the critical trade-offs involved in selecting an appropriate chunking strategy.