Which problem does retrieval-augmented generation address?
Why D is the answer
Answering from content outside the model's training data
RAG retrieves relevant content and passes it to the model.
In detail
- RAG answers from content the model was not trained on.
- The model stays the same.
Use case
An illustrative scenario, not a real incident.
An assistant must answer from this year's policy, which it never saw in training. RAG retrieves it.
See it in different ways
flowchart LR A["Curate content"] --> B["Embed and store"] C["User question"] --> D["Retrieve passages"] B --> D D --> E["Prompt with context"] E --> F["LLM answer"]
sequenceDiagram participant U as Query participant E as Embedding model participant V as Vector database U->>E: Text E-->>U: Vector U->>V: Find the nearest V-->>U: Similar passages
sequenceDiagram participant P0 as User participant P1 as App participant P2 as Document store participant P3 as LLM P0->>P1: Question P1->>P2: Search P1->>P3: Question and passages P3-->>P0: Answer
Key point: RAG brings in outside content.
Why the other options are wrong
- A. Making the model's weights smaller, so they fit on one GPU
- It does not shrink weights.
- B. Reducing the time needed to pretrain from scratch
- It is not about pretraining time.
- C. Removing the need for a tokenizer before inference
- A tokenizer is still needed.