RAG is an engineering problem, not a prompt trick

Teams often treat retrieval-augmented generation as a prompting exercise. In practice, most quality problems trace back to retrieval, not generation.

Chunking strategy decides what the model can even see. Chunks that split a clause across boundaries produce confident answers built on half a sentence.

Retrieval configuration decides what surfaces. Hybrid search, metadata filters and reranking usually beat any amount of prompt tuning.

Evaluation decides whether you actually improved anything. Without a small, honest test set and a scoring method, every change is a matter of opinion.