Retrieve. Learn. Remember.
Three different needs.

RAG for retrieving information

Augmented search, or Retrieval-Augmented Generation, retrieves useful passages and supplies them to the model to prepare an answer. It suits information that you want to keep up to date and tie to sources. Quality depends in particular on the corpus, the search and the way the answer uses the passages.

Example: finding a valid maintenance procedure and citing the relevant section. Versions and viewing rights must also be checked.

Fine-tuning for adapting a behavior

Fine-tuning adjusts an existing model on examples. It can be considered for a classification, a recurring format or a specific behavior when instructions alone are not enough. It does not replace a knowledge base kept up to date.

Example: assigning categories specific to your business to highly repetitive requests. You need usable examples, stable criteria and a comparison on cases outside the training set. The model’s license and the rights to the data are part of the choice.

Memory for keeping a state

An application memory can keep a validated choice, the progress of a case or an authorized preference. It does not necessarily involve changing the model’s weights. You need to specify what is kept, for whom, for how long and how to correct or remove a piece of information.

Example: a case has already been reviewed and is waiting for a document. The system must retrieve that state, not deduce it each time from a full conversation.

These levers can be combined

An assistant can consult up-to-date procedures, produce a consistent business output and keep the state of the case. Start by identifying the flaw: missing information, unstable behavior or lost context. A combination of techniques is only useful if it corrects an observed flaw.

Technical references: RAG in Azure AI Search and model selection and adaptation, OpenAI. These documents shed light on the mechanisms; they are not a measurement of your own use case.