A library consortium asks whether indexing ordinance PDFs for retrieval is a different workload than fine-tuning a generative model on those PDFs. How should the municipal AI architect distinguish them?

The library consortium is asking the municipal AI architect to separate two workloads that happen to touch the same ordinance PDFs. RAG indexing and retrieval prepare documents—chunking, embedding or lexical indexing, and store updates—so queries can look up grounded ordinance text without changing foundation-model weights. Fine-tuning is a training-style workload that updates model parameters using the corpus (or derivatives of it) and therefore changes what the model itself encodes. Saying fine-tuning prepares ordinance chunks for grounded lookup while RAG indexing updates model weights on new amendments reverses the workloads; RAG indexing prepares chunks for grounded lookup, while model weight updates on new amendments come from fine-tuning, not from indexing. Saying RAG indexing and retrieval update model weights on ordinances while fine-tuning builds a vector index for grounded lookup also reverses the workloads; RAG indexing builds the vector index for lookup without changing weights, while fine-tuning updates model weights and does not build a retrieval index. Although RAG indexing and retrieval do keep ordinance embeddings for search, fine-tuning does not rewrite retrieved passages for grounded lookup; rewriting with retrieved context happens at generation time, while fine-tuning updates model weights on the corpus. Exam caveat: some municipal designs combine light fine-tuning with RAG; that hybrid still requires clear separation of weight-update jobs versus index-refresh jobs in capacity planning. Operational check: when ordinances change, decide whether you re-index for retrieval, schedule a fine-tune, or both—and ensure GPU training capacity is not blocked solely because someone labeled every PDF touch as fine-tuning.