Over the past two years, warehouse projects built on large language models have followed a clear pattern: almost all of them start with knowledge Q&A. An operator asks how to pick a certain SKU, and the system finds the answer in manuals, work instructions, and past work orders. The question is not whether to build it, but how. Two technical paths sit on the table — fine-tuning and RAG — and both can look good in a demo. The cost of choosing wrong only shows up six months after go-live.
I have followed three real warehouse knowledge base projects. One was a 3PL customer-service Q&A in East China, launched in early 2024 on fine-tuning. One was an internal operations assistant in a South China e-commerce warehouse, launched in mid-2024 on RAG. The third was a spare-parts lookup system at a North China manufacturer that trialed both approaches in early 2025. All three demoed well. Six months later, their fates had diverged sharply.
The fine-tuning project hit trouble first. Its knowledge base covered more than 40 operating documents, and early accuracy looked fine. The problem was updates. Every time the warehouse changed a process or swapped equipment parameters, the team had to prepare new training data and run another fine-tuning round — two weeks at best, a month at worst. Worse, the model's grasp of new knowledge was unstable. After one update, it began mixing old and new procedures in the same answer: an operator asked about a picking rule and got half the new version and half the old one. The project lead later told me half their effort went into "stopping the model from forgetting things" rather than making the system better.
The RAG project moved more steadily. Knowledge lived as documents in a vector store, so a process change meant swapping a document, effective within minutes. When they installed a new sorting system in late 2024, the updated work instructions were searchable the same day. But RAG had its own weakness: when retrieval missed, the answer could be confidently wrong. Once, an operator asked for the cold-storage temperature alarm threshold, and the system pulled an outdated figure from an expired document, answering with total confidence. The shift nearly followed the wrong standard. The post-mortem was clear — the fault lay not with the model but with knowledge governance. Nobody had cleaned out the expired documents, so retrieval returned garbage.
Fine-Tuning vs RAG: Where the Real Boundary Lies
The boundary between the two paths is not about which is "smarter." It is about the shape of the knowledge and how fast it changes. Fine-tuning bakes knowledge into the model's weights, which suits stable, general, long-lived capabilities: tone, answer format, domain common sense. RAG hangs knowledge outside the model, which suits content that changes often and needs precise citation: operating procedures, equipment parameters, customer contract terms.
Most of a warehouse knowledge base falls into the second category. Processes change, equipment gets replaced, customers add requirements, regulations get updated. If every update demands retraining, maintenance costs quickly outgrow the system's value. That is why my default position is this: RAG should be the default architecture for a warehouse knowledge base, with fine-tuning used only where it earns its place.
The comparison below gives a practical selection framework:
| Dimension | Fine-tuning | RAG |
|---|---|---|
| Cost of updating knowledge | High — retraining required | Low — swap the document |
| Traceability of answers | Poor — no source to cite | Good — points to the source document |
| Hallucination risk | Medium — answers from memory | Depends on retrieval quality, controllable |
| Best fit | Stable general capability, style | Frequently changing specific knowledge |
| Data needed | Hundreds to thousands of quality samples | A few dozen documents to start |
| Typical failure | Old knowledge interferes, catastrophic forgetting | Stale documents pollute retrieval |
"Catastrophic forgetting" deserves a closer look. Fine-tuning rewrites model weights with new data, and pushed too far, it damages capabilities the model already had. The fine-tuning 3PL saw exactly this: after three rounds tuned to new procedures, the model started failing basic arithmetic questions. They had to cut the tuning rounds to a minimum, and the gains shrank with them.
Choosing RAG Is Only the Beginning
Picking RAG does not mean picking correctly — it means picking a starting point that is easier to fix. In a RAG system, roughly 70 percent of success is knowledge governance and 30 percent is technology. Governance means three things: version control for documents, regular cleanup of expired content, and continuous evaluation of retrieval quality.
The South China e-commerce warehouse offers a useful reference. Every document entering the base carries an effective date and an expiry date; expired documents leave the retrieval pool automatically. Each week, the team samples 50 real questions, checks by hand whether the answer cited the right document, and triggers a review if retrieval accuracy falls below a threshold. After six months of this routine, answer accuracy climbed from around 70 percent at launch to a stable 90 percent or better. Architecture sets the ceiling. Governance sets the floor.
Here is a counterintuitive point: when your data is small, do not rush to fine-tune. Many teams start there, believing a model "trained on our data" will understand their business. In practice, fine-tuning on a few hundred samples overfits easily — the model aces questions from its training set and falls apart on phrasings it has never seen. Rather than spending two months assembling fine-tuning data, get RAG running first, let real questions flow in, and decide about fine-tuning once you have enough evidence.
My recommendation is a three-step path. Step one: launch RAG on existing documents, run it for three months, and collect real questions and failure cases. Step two: let the failures decide whether to add fine-tuning — if the failures cluster around tone and format, fine-tuning is worth it; if they cluster around wrong knowledge, fix the knowledge base first. Step three: combine both — fine-tuning for how the system speaks, RAG for what it says.
There is no universal answer to this selection, but there is a universal test: six months from now, is your team using the system, or serving it? If every business change forces the technical team into overtime wrestling with the model, the selection was probably wrong. A warehouse knowledge base earns its keep by changing with the business, not by asking the business to accommodate it.
By Dr. Chen, Supply Chain Researcher







