NCA-GENL practice question 13 of 15

What is 'chunking' in Retrieval-Augmented Generation (RAG)?

Full set: 95 questions

Domain Large Language Models (LLMs)

Answer options
A: Rewrite blocks of text to fill a context window.
B: A method used in RAG to generate random text.
C: A concept in RAG that refers to the training of large language models.
D: A technique used in RAG to split text into meaningful segments. Correct answer
Correct Answer:
D. A technique used in RAG to split text into meaningful segments.
Explanation

In Retrieval-Augmented Generation (RAG), 'chunking' is a critical data preprocessing technique. It involves breaking down large documents from a knowledge base into smaller, semantically coherent segments or 'chunks'. This process is essential for creating effective vector embeddings. By dividing the text, the system can more accurately identify and retrieve the most relevant, specific passages of information in response to a user's query. These targeted chunks are then provided to the Large Language Model (LLM) as context, leading to more accurate and relevant generated answers.

Why Incorrect

A. Chunking is about splitting existing text into segments, not rewriting or generating new text to fill a context window.
B. RAG is designed to generate informed, contextually relevant text based on retrieved data, not random text.
C. Chunking is a data preparation step for the retrieval component of a RAG system, not a method for training the underlying LLM.

References

1. NVIDIA Technical Blog: In the article "What Is Retrieval-Augmented Generation?", it is stated under the "How RAG Works" section: "The first step is called indexing. This is a process for preparing the knowledge base. It involves loading the data, splitting it into chunks, and converting those chunks into numerical representations using an embedding model."

Source: NVIDIA Technical Blog, "What Is Retrieval-Augmented Generation?", August 9, 2023.

2. University Courseware: Stanford University's CS25 course materials on "Transformers United" explain the RAG pipeline. In the section on "Building a Vector Database," the first step described is "Chunking," defined as breaking down documents into smaller pieces to be embedded.

Source: Stanford University, CS25: Transformers United, Fall 2023, Lecture 5 Notes, Section: "Building a Vector Database".

3. Peer-reviewed Academic Publication: The foundational paper on RAG describes retrieving "passages" from a corpus. Modern implementations and analyses of this process explicitly define the division of documents into these passages as "chunking," which is fundamental to the retrieval mechanism's success.

Source: Lewis, P., et al. (2020). "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33, Section 2.1, "Retriever". (This paper establishes the concept of retrieving passages, which is now implemented via chunking).

Scroll to Top