Lesson 05: How RAG relates to training and agents

Training, RAG, and agents often appear together and can blur into making AI smarter. A library assistant makes their roles easier to distinguish.

Does a new rule require retraining?

Training changes model parameters. Fine-tuning is training too and can adapt tasks, style, or outputs. Updating frequently changing facts such as this year’s borrowing limit through training is not necessarily the most convenient maintenance strategy.

RAG finds evidence when answering. Update documents and indexes so retrieval finds B; an unchanged model can then answer three from current context. Uploading a document for reading should not automatically be called training.

Training and RAG can coexist: fine-tuning can adapt output style while RAG supplies current rules. Identify whether the need concerns task behavior or changing knowledge.

Explaining a rule differs from carrying out a task

G answers a two-hour room limit. Booking a room at 14:00 today also requires checking live availability, submitting a reservation, and confirming it. Policy evidence does not establish current availability.

Agents commonly involve model-directed choices, tools, and feedback. A fixed workflow can also book rooms; API calls alone do not make a system an agent. Look for task decisions, execution, and feedback.

An agent can use RAG for rules and then a booking tool. A regular RAG workflow can also use APIs for evidence. They can combine: retrieval-supported answering and broader task execution address different roles.

Separate these three needs

Consider consistent formatting, answers using current borrowing rules, and completing a reservation. Write what you would examine first for each, then explain the differences aloud using the library examples.

One way to decide

Try instructions and examples for formatting before considering fine-tuning. Current rules require maintained evidence and retrieval. Booking needs execution tools and a workflow, potentially an agent. These methods can coexist without all being required for every simple task.

Further reading: Anthropic: Building effective agents

← Previous · Learning plan · Fictional library practice sources · Next →