1Z0-184-25 Question 129
Select 2You are designing a knowledge platform using Retrieval-Augmented Generation (RAG) on Oracle Cloud Infrastructure (OCI) to answer user queries about thousands of internal policy documents stored in OCI Object Storage. You plan to embed these documents and store them in a vector database for effective retrieval. To ensure data accuracy, reduce potential hallucinations, and seamlessly scale as your corpus grows, which two actions should you include in your RAG pipeline implementation?
- A
Use only a large language model fine-tuned on general data without retrieving policy documents, to reduce complexity.
- B
Retrieve relevant document segments from the vector store and supply them as context to the model before generation.
- C
Store the entire policy corpus in a traditional relational database to avoid the overhead of storing embeddings separately.
- D
Continuously update your embeddings and vector indexes as new documents are added or existing documents change.
Show answer and explanation
Correct answers: B, D
Explanation
Retrieval-Augmented Generation (RAG) improves answer correctness by grounding language model responses in specific, relevant documents. In OCI, storing embeddings in a vector-indexed system and regularly updating them as your corpus evolves is key. By retrieving the correct segments and providing them as context to the language model, you reduce hallucinations and increase factual accuracy. For more details, refer to OCI best practices for AI and machine learning deployments, and Oracle� Documentation on generative AI and data science pipelines.
- A. Incorrect.
Incorrect. Relying solely on a model fine-tuned on general data without retrieval greatly increases the risk of hallucination and reduces accuracy for domain-specific queries. RAG specifically aims to ground answers in relevant documents, so skipping the retrieval step undermines the entire approach.
- B. Correct.
Correct. A core principle of RAG is to retrieve the most relevant snippets from your vector store and feed them to the language model. Providing domain-specific context at generation time is crucial for accurate, grounded responses.
- C. Incorrect.
Incorrect. A traditional relational database may not be suitable for fast similarity search needed by RAG. Storing vector embeddings in a specialized vector database or search service is recommended to efficiently retrieve the most relevant documents.
- D. Correct.
Correct. Policy documents will change over time, and new ones will be added. Regularly updating embeddings and reindexing ensures the retrieval step stays accurate. This helps maintain the reliability of the generated responses.