Fine-tuning and retrieval-augmented generation

At a glance

Fine-tuning changes

Model behavior through training

Retrieval supplies

External context at request time

Overview

Fine-tuning changes a model using training examples, while retrieval-augmented generation supplies source material during a request. They address different parts of an AI application's behavior and can be combined, but should be evaluated against a clearly defined task and failure pattern.

What differs

Fine-tuning uses examples to adapt a model's behavior. Retrieval selects information from an external collection for the current request, making the source collection a separate part of the application.

Choose by the task

Diagnose whether the problem concerns behavior, missing context, or both. Evaluate representative inputs and desired outputs before investing in a technique, and compare the result with a simpler baseline.

Common confusion

Retrieval does not guarantee faithful use of evidence, and fine-tuning does not automatically keep knowledge current. Each approach still needs evaluation and a process for maintaining its data and configuration.

Sources and review

MOOR's explanatory text is supported by the following source links.

1. Model optimization — OpenAI
2. Retrieval-augmented generation — Cohere