Exercise 3: Comparing Three Chunking Methods¶

Chunking means cutting a long document into a few small boxes. Boxes that are too big bury the detail; boxes that are too small cut the meaning in half.

📌 Learning goals¶

Name the three cutting methods: fixed-length, paragraph-based, and heading-aware.
Know that chunk size and overlap change search results.
Fairly compare all three strategies on the same document and the same set of questions.
Block a bad overlap value that would hang the program.

🔑 Core terms¶

Core term	Plain meaning
Chunk	A small piece cut out of a long document
Chunk size	The maximum size of each piece
Overlap	Text repeated between two adjacent chunks, so a sentence doesn't get cut right in half
Retrieval	Using a question to pull back the most useful passages

📚 Required reading and learning resources¶

★★★★★ LangChain text splitters official concepts page: start with the simple cutting methods.
★★★★★ LlamaIndex Semantic Splitter official API: go deeper once you need semantic breakpoints.
★★★★☆ Unstructured official docs: an entry point for parsing complex formats like PDF, DOCX, and HTML.
★★★★☆ datawhalechina/hello-agents: a fuller course on RAG chunking.

_{Data verified: 2026-08-30 UTC.}

▶️ Run Path A first (local, provider-independent, free)¶

pip install -r requirements.txt
python starter.py

The first run may download a local embedding model; the API cost is $0.

Path B (preview: how this step feeds into Claude later)

The chunking step only decides how to split a document — it doesn't care which model generates the final answer later. Running starter_anthropic.py here just shows you the same chunking code that will plug in, unchanged, right before the Claude generation step in Exercise 4. This step makes no API call, so the cost is $0.

python starter_anthropic.py

Total Stage 06 budget: Running all five Path A exercises keeps API fees at $0 (downloads, disk space, and electricity excluded). Optional cloud paths are billed from actual embedding, input, and output token usage; set a small account cap and stop after one successful run.

python test.py
python test_anthropic.py

The tests only check the chunking logic — no model download.

Three cutting methods¶

Strategy	How it cuts	Good for
Fixed-length	Cuts every N characters	Logs or chat transcripts with no clear format
Paragraph-based	Cuts on blank lines	Articles with tidy paragraphs
Heading-aware	Cuts on `#`, `##` headings	READMEs, wikis, specs

fixed = chunk_fixed(text, chunk_size=200, overlap=40)
paragraphs = chunk_paragraphs(text)
headings = chunk_headings(text)

overlap must satisfy:

0 <= overlap < chunk_size

If overlap is equal to or greater than chunk_size, the next chunk never moves forward. The program now raises ValueError right away instead of letting it become an infinite loop.

How to judge which is better¶

Don't just count chunks, and don't judge from one pretty example. Prepare real questions and check whether the passage holding the correct answer actually made it into the top-k.

What to check	Ask yourself
Recall	Was the correct passage retrieved at all?
Precision	Of what was retrieved, how much is actually useful?
Completeness	Did the sentences the answer needs get split across two chunks?
Cost	How much do embedding and storage grow as chunks multiply?

Common pitfalls and advanced practice

- Chunks too big: too many topics crammed into one piece, so the vector gets blurry. - Chunks too small: the context an answer needs may end up scattered across different pieces. - A PDF isn't plain text: handle columns, tables, headers, and footers before chunking. - Don't hard-cut CJK text by byte count; use Python strings, a tokenizer, or a structured parser. - Advanced approach: cut into large sections by heading first, then cut those sections into smaller fixed-length pieces.

Next: connect chunking, embedding, search, and answering into Exercise 4: Full RAG.