AWS Certified Generative AI – Professional (AIP-C01) — Question 57
A legal research company has a Retrieval Augmented Generation (RAG) application that uses Amazon Bedrock and Amazon OpenSearch Service. The application stores 768-dimensional vector embeddings for 15 million legal documents, including statutes, court rulings, and case summaries.
The company's current chunking strategy segments text into fixed-length blocks of 500 tokens. The current chunking strategy often splits contextually linked information such as legal arguments, court opinions, or statute references across separate chunks. Researchers report that generated outputs frequently omit key context or cite outdated legal information.
Recent application logs show a 40% increase in response times. The p95 latency metric exceeds 2 seconds. The company expects storage needs for the application to grow from 90 GB to 360 GB within a year.
The company needs a solution to improve retrieval relevance and system performance at scale.
Which solution will meet these requirements?
Answer options
- A. Increase the embedding vector dimensionality from 768 to 4,096 without changing the existing chunking or pre-processing strategy.
- B. Replace dynamic retrieval with static, pre-written summaries that are stored in Amazon S3. Use Amazon CloudFront to serve the summaries to reduce compute demand and improve predictability.
- C. Update the chunking strategy to use semantic boundaries such as complete legal arguments, clauses, or sections rather than fixed token limits. Regenerate vector embeddings to align with the new chunk structure.
- D. Migrate from OpenSearch Service to Amazon DynamoDB. Implement keyword-based indexes to enable faster lookups for legal concepts.
Correct answer: C
Explanation
Option C is correct because updating the chunking strategy to use semantic boundaries will ensure that contextually linked information remains intact, improving the relevance of retrievals. This will also allow for more accurate vector embeddings that align with the new structure. Options A and B do not address the issue of context loss and retrieval accuracy effectively, while Option D suggests a database change that would not solve the core problem of contextual segmentation.