RAG First, Fine-Tune Last: The Boring Answer Wins Again
A practitioner's guide and an academic survey landed two days apart, both with the same advice: before you fine-tune anything, fix your retrieval. RAG has quietly graduated to infrastructure.
A practitioner's field guide and a freshly published ACM survey landed within two days of each other this week, both pointing at the same boring truth: in 2026, the answer is still retrieval-augmented generation before fine-tuning. This isn't a fashion swing — it's a maturity signal that the knowledge integration playbook has finally consolidated.
A practitioner's gut and an academic stamp
The hjLabs 2026 best-practices guide opens with what amounts to a pre-commit warning: try RAG and prompt engineering before fine-tuning, because (per the author's field experience as of May 2026) roughly 80% of fine-tuning requests turn out to be retrieval problems in disguise. That 80% figure is anecdotal practitioner observation rather than a benchmark, but it tracks with what most teams discover the second time they spend a weekend on a QLoRA run and realize their chunking was the bottleneck all along.
Two days later, ACM Computing Surveys published A Survey on Retrieval-Augmented Text Generation for Large Language Models (15 May 2026). Surveys arrive when a field is stable enough to map, not when it's still being invented. A survey-level treatment from a venue as conservative as ACM signals that RAG has crossed from research pattern to infrastructure pattern.
What it means depending on your seat
For builders, this is permission to stop apologizing for the unsexy choice. If your roadmap has a fine-tuning project that's been delayed 3 quarters, kill it and double your retrieval budget instead. The hjLabs piece names QLoRA on a single A100 80GB as the sensible default when you do need to fine-tune — but the path there is paved with the failed RAG iterations that ruled it out first.
For buyers and PMs, the implication is sharper: vendor pitches leading with 'we fine-tune a custom model for you' deserve more scrutiny than ones leading with 'we structured your knowledge base for retrieval.' The former is harder to maintain, harder to evaluate, and frequently doesn't beat a carefully built RAG pipeline.
For content owners, the survey's existence means your retrievability is now an evergreen asset. The teams winning at RAG aren't the ones with the cleverest rerankers — they're the ones whose source material is structured, dated, and chunked in a way the retrieval layer can actually use.
The bottom line: If you're about to fine-tune, spend one more sprint on retrieval first — the survey just shipped, the field has consolidated, and your QLoRA budget is almost certainly better spent on chunking.