Databricks Generative AI Engineer Associate Question 259
Select 3You are building a Retrieval-Augmented Generation (RAG) application that relies on a large external dataset as its primary data source. During testing, you notice that some of the text in the dataset contains offensive or biased language, which may negatively affect the quality and ethics of the generated responses. What is an effective alternative approach to mitigate this issue?
- A
Introduce a data preprocessing pipeline to clean and filter problematic text before it is indexed for retrieval.
- B
Use a post-generation filter that identifies and removes offensive or biased output in the generated responses.
- C
Switch to a smaller dataset that is manually curated for high-quality, unbiased content.
- D
Fine-tune the generative model using a dataset specifically designed to reduce bias and offensive language.
- E
Rely solely on user feedback to identify and mitigate problematic responses after deployment.
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 data source ensures that harmful text does not enter the retrieval pipeline, while a post-generation filter safeguards the output. Fine-tuning the model further reduces the likelihood of generating biased or offensive responses. Solely relying on user feedback or limiting the dataset's scope are less effective or impractical solutions for robust and scalable RAG systems.
- A. Correct.
Introducing a data preprocessing pipeline to clean and filter problematic text before indexing ensures that the retrieval stage avoids exposing the model to harmful content. This step helps mitigate issues early in the pipeline.
- B. Correct.
Using a post-generation filter adds a layer of protection by ensuring that even if problematic text is retrieved, the final generated output is filtered or altered to remove offensive or biased language.
- C. Incorrect.
Switching to a smaller manually curated dataset may reduce bias but can significantly limit the scope and relevance of the retrieved information, which could negatively impact the RAG application's performance.
- D. Correct.
Fine-tuning the generative model using a dataset designed to reduce bias and offensive language helps the model learn to generate better responses, complementing other mitigation strategies.
- E. Incorrect.
Relying solely on user feedback is reactive and does not provide a proactive solution to the issue, leading to potential ethical and operational challenges before the problem is identified.