1Z0-184-25 Question 121
Single answerYou are designing a Retrieval-Augmented Generation (RAG) application on Oracle Cloud Infrastructure (OCI) to provide real-time answers from thousands of product manuals. The PDFs are stored in Oracle Object Storage, and you need to generate vector embeddings of their text content for efficient similarity-based retrieval. The application must handle a large number of concurrent queries and deliver minimal response times. Which approach best meets these requirements?
- A
Use Oracle Functions to generate embeddings on-the-fly for each query directly from Object Storage, caching intermediate vectors in function memory.
- B
Use Oracle Container Engine for Kubernetes to run a pipeline that preprocesses PDFs into embeddings, stores the vectors in an Oracle Database with vector search capabilities, and retrieves them for a generative AI model at query time.
- C
Maintain an in-memory cache of all embeddings on a single OCI Compute instance, refreshed nightly from Object Storage, to serve similarity search queries.
- D
Deploy the AI model in Oracle Data Science and store only text chunks in an Oracle Object Storage bucket, relying on file-based searches for retrieval.
Show answer and explanation
Correct answer: B
Explanation
The recommended design for a RAG application in OCI includes separating the embedding generation pipeline (often run on Oracle Container Engine for Kubernetes or another scalable service) from the vector database layer (for example, Oracle Database with vector search support). This ensures that embeddings are generated once, then stored and indexed efficiently for subsequent similarity lookups, minimizing both latency and infrastructure overhead. Refer to Oracle� official documentation on Vector Search in Oracle Database 23c and best practices for containerized workloads on OCI for detailed implementation guidance.
- A. Incorrect.
Option 1: Incorrect � While Oracle Functions is useful for lightweight, event-driven tasks, generating embeddings on-the-fly for each query will lead to high latency and high operational overhead. Ephemeral function memory is also not suitable for storing large vector datasets.
- B. Correct.
Option 2: Correct � By using a container-based embedding pipeline, you can preprocess documents in bulk. Storing the resulting embeddings in an Oracle Database with vector search capabilities enables efficient similarity searches and reduces query response times for large workloads. This architectural approach is both scalable and cost-effective.
- C. Incorrect.
Option 3: Incorrect � Keeping all embeddings in a single compute instance's memory is risky and not easily scalable. You may run out of memory as the dataset grows, and refreshing the entire cache nightly can cause availability or performance issues.
- D. Incorrect.
Option 4: Incorrect � Storing only text chunks or embeddings as files in Object Storage and relying on ad-hoc file searches is not efficient for low-latency vector similarity queries. You would need a robust, searchable store to effectively retrieve relevant content.