Our documentation changes every week. Why not just fine...

Relying on weekly fine-tuning for rapidly changing documentation is flawed both mathematically and computationally compared to RAG:

1. Catastrophic Forgetting: Standard fine-tuning loss functions do not include terms to preserve existing weights. Forgetting follows a quadratic function of parameter updates $\frac{1}{2} \Delta^T F \Delta$ (where $F$ is the Fisher information matrix), meaning parameter drift destroys old knowledge quadratically with respect to step size, whereas newly acquired domain knowledge grows only linearly.
2. FLOP Inefficiency: Updating model weights costs roughly $6N_{\text{gen}}$ FLOPs per token (for a generator model of size $N_{\text{gen}}$). Updating a vector index using an embedding model costs roughly $2N_{\text{emb}}$ FLOPs per token. For a 27B parameter generator and a 110M parameter encoder, fine-tuning is approximately 736 times more expensive per token than indexing, regardless of overall corpus size.

Fine-tuning is intended for teaching style, output structure, and vocabulary adaptations, whereas retrieval remains the correct, cost-effective tool for updating factual and rapidly changing knowledge bases.