RAG vs Fine-Tuning¶
When to retrieve external knowledge vs when to adapt the model itself.
Dimensions of Comparison¶
| Dimension | RAG | Fine-Tuning |
|---|---|---|
| Knowledge updates | Immediate (update document store) | Requires retraining |
| Traceability | High (can cite source documents) | Low (knowledge in weights) |
| Cost | Per-query retrieval cost | Upfront training cost |
| Domain vocabulary | Via retrieved docs | Learned into weights |
| Hallucination | Reduced when grounded | May reduce or shift |
| Independence | Same reasoning, different knowledge | Partial decorrelation |
| Air-gap compatibility | Both work locally | Both work locally |
When to Use RAG¶
- Factual knowledge that changes (procedures, Tech Specs, operating experience)
- Need to cite specific source documents
- Knowledge must be auditable and updatable
- Primary approach for nuclear document grounding (Report 1 §8.1)
When to Fine-Tune¶
- Consistent domain vocabulary and response style
- Task-specific format adaptation (structured output formats)
- Performance on specific task types (Report 6 §2)
- Combined with RAG: fine-tune for style, RAG for facts
Nuclear Recommendation¶
RAG is the primary knowledge grounding mechanism for nuclear applications. Fine-tuning is supplementary — useful for adapting response format and domain vocabulary but insufficient alone for factual grounding. Knowledge graph constraints provide an additional verification layer that neither RAG nor fine-tuning supplies.
Fine-tuned variants of the same base model retain shared pre-training biases — they provide partial decorrelation, not full epistemic-independence.