Course contentsShow
← Machine Learning and Deep Learning
Lesson 1995 of 3,53842. Retrieval-Augmented Generation: FoundationsFree lesson

Multi-Representation Chunking

Creating multiple chunk versions or summaries of the same content to improve retrieval from different query types.

Multi-Representation Chunking

What you'll learn: How to create multiple versions of the same content chunk to match different types of user queries and improve RAG retrieval accuracy.

The Core Problem

Imagine a user asking "What's the main idea of this paper?" versus "What neural network architecture was used in experiment 3?" Both questions might relate to the same document section, but they're looking for different types of information—one wants a high-level summary, the other wants specific technical details.

Traditional chunking creates one representation per text segment. If that representation doesn't match the query's style or abstraction level, retrieval fails even when the content is relevant.

What Is Multi-Representation Chunking?

Multi-representation chunking stores multiple versions of the same underlying content:

  • Original chunk: The raw text segment
  • Summary version: A condensed, high-level distillation
  • Question-answer pairs: Hypothetical questions this chunk could answer
  • Keyword-enriched version: The chunk with extracted key terms highlighted

During indexing, you embed all these representations separately but link them back to the same source chunk. When a user query comes in, it might match the summary (for conceptual questions) or the original text (for detail-seeking queries).

How It Works

  1. Chunk your document using any strategy (sentence-based, semantic, etc.)
  2. Generate alternate representations for each chunk—use an LLM to create summaries or hypothetical questions
  3. Embed all versions and store them in your vector database with metadata pointing to the original chunk
  4. At retrieval time, the query matches whichever representation is most similar
  5. Return the original chunk for context generation

Key Takeaway

Multi-representation chunking improves retrieval robustness by indexing the same content at different abstraction levels and formats, ensuring diverse query types can successfully find relevant information.