Databricks Generative AI Engineer Associate Question 258
Single answerYou are designing a Retrieval-Augmented Generation (RAG) application that relies on a large textual dataset as its knowledge base. During testing, you notice the presence of problematic text, such as biased, offensive, or irrelevant content, which could negatively impact the responses generated by the model. What is the best alternative to mitigate this issue while maintaining the quality of the knowledge base?
- A
Implement a pre-processing pipeline to filter and clean the problematic text from the source dataset
- B
Reduce the size of the dataset by randomly removing a significant portion of the documents
- C
Use a post-processing step to remove or rephrase problematic content in the model's output
- D
Retrain the language model on additional curated datasets to reduce sensitivity to problematic content
Show answer and explanation
Correct answer: A
Explanation
The best approach to mitigate problematic text in the data source for a RAG application is to implement a pre-processing pipeline. This method directly addresses the issue at the source and ensures the quality and reliability of the knowledge base feeding the RAG system. While other options may offer partial solutions, they are either less effective or introduce additional challenges.
- A. Correct.
Implementing a pre-processing pipeline addresses the issue directly at the data source level by removing or cleaning problematic text before it reaches the RAG application. This ensures the quality of the knowledge base and avoids propagating harmful or irrelevant content.
- B. Incorrect.
Randomly removing documents from the dataset does not specifically target problematic text and may result in the loss of valuable information, thereby reducing the effectiveness of the RAG system.
- C. Incorrect.
Using a post-processing step only addresses problematic content after it has been generated by the model, which is less effective than addressing the issue at the source. It also adds complexity to the application and may not fully mitigate the problem.
- D. Incorrect.
Retraining the language model can improve its performance in some scenarios, but it does not directly solve the issue of problematic text in the data source for a RAG application. Additionally, retraining can be resource-intensive and time-consuming.