Quest 20 - RAG Chunking Strategist

Quest 20: RAG Chunking Strategist

medium 25 minutes

---

🎯 Learning Objectives

• ✅ Understand why document chunking is critical for RAG retrieval quality
• ✅ Implement paragraph, sentence, and word-based chunking strategies
• ✅ Apply overlap between chunks to preserve context across boundaries
• ✅ Recognize the tradeoff between chunk size and retrieval precision

---

📖 Concept: Document Chunking

RAG (Retrieval-Augmented Generation) ต้องการแบ่งเอกสารขนาดใหญ่เป็นชิ้นเล็ก ๆ (chunks) เพื่อค้นหาข้อมูลที่เกี่ยวข้อง แต่ วิธีแบ่งส่งผลโดยตรงต่อคุณภาพการค้นหา

แบ่งเล็กเกินไป → 丢失 context สำคัญ แบ่งใหญ่เกินไป → relevance score จางลงเพราะมี noise เยอะ

Chunking คือศิลปะแห่ง balance — เหมือน martial arts ที่ต้องหาจุดสมดุลระหว่าง speed และ power

---

⚙️ How It Works

Chunking Strategies

Strategy        | Boundary       | Best For
────────────────|────────────────|─────────────────────
paragraph       | \n\n           | Well-structured docs
sentence        | . ? !          | Prose, articles
words           | whitespace     | Technical docs, code

The Chunking Workflow

1. Receive text and options
         ↓
2. Split by chosen boundary (paragraph/sentence/words)
         ↓
3. Merge small pieces up to maxChunkSize
         ↓
4. Apply overlap between adjacent chunks
         ↓
5. Return array of text chunks

Why Overlap Matters

Without overlap, the boundary between chunks loses context:

Chunk 1: "The cat sat on the"
Chunk 2: "mat and looked outside"
# "sat on the mat" is split across chunks!

With overlap:

Chunk 1: "The cat sat on the"
Chunk 2: "on the mat and looked outside"
# Context preserved!

---

💡 Example: Implementing Chunking

Step 1: Split by paragraph

function chunkDocument(text, options) {
  const { maxChunkSize = 500, overlap = 0, splitBy = 'paragraph' } = options;

  let pieces;
  switch (splitBy) {
    case 'paragraph': pieces = text.split(/\n\n+/); break;
    case 'sentence':  pieces = text.split(/(?<=[.!?])\s+/); break;
    case 'words':     pieces = text.split(/\s+/); break;
  }

  return mergeAndOverlap(pieces, maxChunkSize, overlap);
}

Step 2: Merge small pieces

function mergeAndOverlap(pieces, maxSize, overlap) {
  const chunks = [];
  let current = '';

  for (const piece of pieces) {
    if (current.length + piece.length > maxSize && current.length > 0) {
      chunks.push(current.trim());
      // Apply overlap from end of current chunk
      current = overlap > 0 ? current.slice(-overlap) + ' ' + piece : piece;
    } else {
      current += (current ? ' ' : '') + piece;
    }
  }
  if (current.trim()) chunks.push(current.trim());
  return chunks;
}

Step 3: Verify

node test.js
# Test: splits into multiple chunks
# Test: chunks respect max size
# Test: overlap creates shared content
# Test: word split respects boundaries

---

⚠️ Common Mistakes

Mistake 1: Splitting mid-word

“Hello world” → [“Hello”, “wor”, “ld”] → Always split at natural boundaries (paragraphs, sentences, or words).

Mistake 2: No overlap

Chunks have zero shared context at boundaries → Add overlap (5-20% of chunk size) to preserve context.

Mistake 3: Chunks too large

One chunk contains the entire document → Set reasonable maxChunkSize (200-500 chars for most use cases).

Mistake 4: Ignoring edge cases

Empty text, single-word text, text shorter than maxChunkSize → Always handle gracefully — return at least one chunk.

---

📝 Knowledge Check

📝 Knowledge Check

Q1:Why is document chunking important for RAG systems?

It reduces storage costsToo small loses context, too large dilutes relevance — chunking balances bothIt's required by the embedding modelIt makes documents shorter for faster reading

Q2:What is the purpose of overlap between chunks?

To increase the total number of chunksTo preserve context that would otherwise be lost at chunk boundariesTo make chunks biggerTo reduce processing time

Q3:What is a common edge case when splitting by words?

Splitting in the middle of a wordCreating too many chunksLosing punctuationChanging the case of words

---

🏋️ Quest: RAG Chunking Strategist

Now it’s time to build a document chunking system!

1. Download ไฟล์เริ่มต้นของ quest:

   npx bluebeltdojo download quest-20-rag-chunking
   cd quest-20-rag-chunking

2. เปิด problem.js ใน editor ของคุณพร้อมความช่วยเหลือของ AI

3. Implement chunkDocument(text, options) — รองรับ paragraph, sentence, word splitting

4. ทดสอบ edge cases: empty text, overlap, max size, boundary respect

5. ตรวจสอบ solution ของคุณ:

   node test.js

6. When all tests pass, submit your solution:

   npx bluebeltdojo submit

💡 Tip: Chunking strategy กำหนดคุณارة RAG ทั้งหมด — แบ่งดี retrieval ดี, แบ่งไม่ดี retrieval พัง

---

คำใบ้

• อ่าน instructions ใน problem.js อย่างละเอียด
• ตรวจสอบ: chunks ต้อง respect boundaries, มี overlap, ไม่เกิน maxChunkSize
• ถ้าติดขัด ลองอ่าน “Common Mistakes” อีกครั้ง — อย่าดู solution โดยตรง