LLM Training, Fine-Tuning, RAG, and Agents: Source Notes

Source Map

Vendor explainers were used only for orientation and were not treated as durable definitions unless they matched the official docs above.

Pretraining and fine-tuning both change model weights, but they differ in scope, cost, and retraining cadence.
Prompting changes runtime context, not weights. Long or highly constrained instructions become brittle more quickly than narrower prompts.
RAG injects retrieved documents into context, so it is suitable for fresh information and internal knowledge, but retrieval misses and bad source quality still leak into the answer.
Tool use gives the model external actions such as calculation, lookup, database operations, or business API calls. A single call is different from a multi-step autonomous loop.
Agents are control loops that iterate through planning, execution, and verification. They are not synonymous with workflows.
MCP is an integration protocol, not a training method. It complements RAG and tool use rather than replacing them.

Vendor marketing claims were downgraded when they lacked evaluation details or stable definitions.
Detailed implementation examples were omitted because the report is a comparison and decision guide, not an engineering cookbook.
Claims that RAG guarantees correctness or that fine-tuning can absorb constantly changing facts were rejected.

The actual lift from fine-tuning depends on the dataset and the evaluation metric.
RAG quality depends on chunking, metadata, retrieval quality, and access control.
Agent safety depends on permissions, sandboxing, auditability, and rollback paths.
MCP’s practical value depends on how much connector sprawl it can reduce in a real deployment.