RAG systems are often judged by the answer the LLM produces. But many of the hardest failures happen before generation even begins.

If the relevant evidence was never retrieved, was ranked too low, was fragmented across poor chunks, or was drowned out by irrelevant context, even the best model cannot produce a reliable answer.

This session looks at RAG as a retrieval engineering problem rather than simply an LLM integration pattern. We will compare lexical, semantic and hybrid retrieval approaches and examine how chunking, metadata, query formulation, ranking and reranking affect the evidence that reaches the model.

Using realistic failure cases, we will trace a RAG question from user query to retrieval to final generation and identify where quality is actually being lost. We will also explore practical ways to evaluate retrieval independently from the final LLM response.

The goal is to help developers stop asking only "Did the LLM answer correctly?" and start asking the more useful question: "Did we retrieve the right evidence in the first place?"