1Z0-184-25 Question 128
Single answerYour team is building a generative AI-based question-and-answer chatbot about your company� product documentation, which is stored in Oracle Cloud Infrastructure. You want to ensure the chatbot can accurately reference the most relevant information from your official knowledge base and minimize hallucinations. Which approach best achieves this goal?
- A
Store the complete set of product documentation in a single large text file and feed it into the language model once at initialization.
- B
Implement a retrieval-augmented generation pipeline where documentation is split into chunks, indexed in a vector store, and retrieved on demand during answer generation.
- C
Use only the pre-trained language model without external references, relying on its internal training data for all domain-specific queries.
- D
Update the language model daily by fine-tuning it with the latest product documents, removing any retrieval component.
Show answer and explanation
Correct answer: B
Explanation
Retrieval-augmented generation (RAG) is crucial when dealing with constantly evolving knowledge bases and large sets of documents. In OCI environments, best practices recommend chunking documents, creating embeddings, and storing them in a specialized vector store or search service. The language model then pulls only the most relevant chunks during generation, significantly reducing the likelihood of hallucinations and ensuring accurate, up-to-date answers. For more details, refer to Oracle� documentation on implementing AI services and best practices for retrieval-augmented generation in OCI.
- A. Incorrect.
Option 1 is incorrect because feeding an entire large text file at initialization can lead to context overflow and inefficiency. It does not address real-time retrieval of targeted information, thus increasing the risk of inaccurate responses.
- B. Correct.
Option 2 is correct. A retrieval-augmented generation pipeline splits textual data into manageable chunks, stores embeddings in a vector index (e.g., in OCI), and retrieves context in real-time based on user queries. This significantly reduces hallucinations by providing relevant details from the knowledge base as the model generates answers.
- C. Incorrect.
Option 3 is incorrect because relying on a pre-trained model without referencing external documents can lead to outdated or irrelevant responses, especially if the domain-specific content has changed since the model was trained.
- D. Incorrect.
Option 4 is incorrect because continuously re-fine-tuning the model for each knowledge update is costly and time-consuming. A retrieval component (RAG) allows for quicker updates by simply adding new data to the index without retraining the entire model.