問題文
An agent needs to answer from 3 million pages of documentation. Which design is viable?
選択肢
- Fine-tuning the model on the entire corpus so it memorizes the content
- A model with the largest available context window, into which the full corpus is loaded
- A knowledge source with retrieval, so only the passages relevant to each question enter the request
- Summarizing the corpus into a single 10-page document and putting it in the system prompt, which keeps every request small enough that no retrieval step is ever needed