A recent Towards Data Science essay warns that enterprises waste months fine‑tuning Retrieval‑Augmented Generation (RAG) pipelines with traditional ML toolkits. The author describes a team that spent six months running five Optuna sweeps, adding a custom reranker and fine‑tuning embeddings, yet production accuracy never budged because the parser was broken. The piece labels this the most costly misconception in enterprise RAG.

The article argues that hyperparameter sweeps, train‑test splits, and explainability frameworks belong to model training, not to search‑system assembly. In RAG the answer either exists verbatim in a document or it does not, making performance binary per query. Mis‑treating chunk size, top‑k or similarity thresholds as learnable parameters yields marginal test‑set gains that disappear in production.

Instead than ML optimizations, teams should invest in the four engineering bricks the author outlines: parsing, question parsing, retrieval and generation. Improving document parsing, crafting robust question prompts, and refining index structures directly significantly raise hit rates, while the underlying embedding model remains a consumed service. The piece ends by urging engineers to treat RAG as a search pipeline, not a trainable model.