กลยุทธ์การแบ่งข้อความตามบริบท

In Retrieval Augmented Generation (RAG), the quality of your retrieved information directly impacts the LLM's response. If the chunks of text you feed into your system are poorly structured, the LLM might miss crucial context.

Think of it like trying to read a book where every other sentence is on a different page. It would be hard to understand the story!

Previously, we touched upon basic text chunking. Often, this involves splitting text into fixed-size segments.

However, fixed-size chunks can cut sentences or paragraphs in half, separating related ideas. This makes it difficult for the retrieval system to find all the necessary information for a query.

• Lost Meaning: Half a sentence often loses its original meaning.
• Incomplete Information: The LLM gets fragments, not full ideas.

Context-aware chunking is a smart way to split your text. Instead of just chopping text at arbitrary lengths, it tries to preserve the natural flow and meaning of the content.

The goal is to keep semantically related pieces of text together within the same chunk. This ensures that when a chunk is retrieved, it provides a complete and coherent piece of information.

A simple yet effective context-aware technique is overlapping chunks. When you split your document, each chunk shares a small portion of text with the previous and next chunks.

This overlap acts as a bridge, ensuring that if an important concept spans a chunk boundary, both chunks will contain enough information to maintain the context.

• Chunk 1: "...the quick brown fox jumped..."
• Chunk 2: "...fox jumped over the lazy dog..."

Notice "fox jumped" is in both, preserving flow.

Another powerful strategy is to split text based on sentence boundaries. This ensures that no sentence is ever broken across two chunks.

Since sentences typically represent complete thoughts, keeping them intact significantly improves the semantic coherence of each chunk, leading to more accurate retrieval.

Let's see a basic Python example of how you might split text into sentences. This helps keep complete thoughts together.

The Recursive Character Text Splitter is a more advanced and widely used technique. It attempts to split text using a list of separators, trying them in order of preference.

If splitting by the first separator (e.g., double newline for paragraphs) results in chunks that are still too large, it recursively tries the next separator (e.g., single newline for lines), and so on.

Imagine you have a long document:

1. It first tries to split by "\n\n" (paragraph breaks).
2. If any resulting chunk is still too big, it takes that big chunk and tries to split it by "\n" (line breaks).
3. If chunks are still too big, it might try " " (spaces for words), or even "" (individual characters) as a last resort.

This hierarchical approach prioritizes preserving larger semantic units before breaking them down further.

The effectiveness of a recursive splitter heavily depends on the list and order of separators you provide. Common separators include:

• "\n\n": For paragraph breaks (strong semantic boundary).
• "\n": For line breaks.
• " ": For word breaks.
• "": For character breaks (the ultimate fallback).

By defining these, you guide the splitter to maintain logical text structures.

Today, we explored how context-aware chunking significantly improves RAG system performance. We moved beyond simple fixed-size splits to methods that respect the natural structure and meaning of text.

• Overlapping chunks bridge information across boundaries.
• Sentence-based chunking keeps complete thoughts intact.
• Recursive character splitting uses a hierarchy of separators to intelligently break down documents.

By implementing these strategies, you ensure that your RAG system retrieves more relevant and coherent information, leading to better LLM responses.