Databricks Generative AI Engineer Associate Question 262
Select 3You are working on a Retrieval-Augmented Generation (RAG) application that relies on a document store as its data source. During testing, your team discovers that some documents contain problematic text, such as offensive language or biased information, which negatively impacts the application's outputs. Which of the following approaches could serve as an alternative to mitigate this issue effectively?
- A
Preprocess the document store by applying a content filter to remove problematic text before indexing.
- B
Implement a real-time text moderation layer during retrieval to filter out problematic text on the fly.
- C
Replace the RAG application with a rule-based chatbot that does not require a document store.
- D
Add a fine-tuned language model to detect and remove problematic content during the generation stage.
- E
Manually curate and review the entire document store to ensure all problematic text is removed.
Show answer and explanation
Correct answers: A, B, D
Explanation
Mitigating problematic text in a RAG application requires a combination of proactive and reactive strategies. Preprocessing the document store ensures cleaner data for indexing, while real-time moderation provides flexibility to address issues dynamically. Adding a fine-tuned language model during generation helps catch any residual issues, offering a robust multi-layered solution. Manual curation and switching to rule-based systems are impractical or counterproductive for scalable AI applications.
- A. Correct.
Preprocessing the document store with a content filter is an effective strategy to mitigate problematic text before it enters the RAG pipeline, ensuring cleaner data for retrieval and generation.
- B. Correct.
Implementing a real-time text moderation layer allows problematic content to be detected and filtered dynamically, adding flexibility for handling new issues as they arise.
- C. Incorrect.
Replacing the RAG application with a rule-based chatbot significantly limits the application's capabilities and is not a practical alternative to mitigate problematic text.
- D. Correct.
Adding a fine-tuned language model for detecting and removing problematic content during the generation stage is a valid approach. It ensures the final generated output aligns with content moderation policies.
- E. Incorrect.
Manually curating the document store is time-intensive and impractical at scale, especially for large datasets in enterprise applications.