Token-based splitting in LangChain

Introduction

Token-based splitting helps break text into smaller parts based on tokens, which are pieces of words or characters. This makes it easier to handle large texts in manageable chunks.

When you want to process long documents without losing context.

When you need to feed text into language models that have token limits.

When you want to split text more precisely than by sentences or paragraphs.

When you want to keep track of token counts for cost or performance reasons.

Syntax

LangChain

from langchain.text_splitter import TokenTextSplitter

splitter = TokenTextSplitter(chunk_size=100, chunk_overlap=20)
chunks = splitter.split_text(text)

chunk_size sets how many tokens each chunk should have.

chunk_overlap sets how many tokens overlap between chunks to keep context.

Examples

This splits the text into chunks of 50 tokens each, with 10 tokens overlapping between chunks.

LangChain

from langchain.text_splitter import TokenTextSplitter

splitter = TokenTextSplitter(chunk_size=50, chunk_overlap=10)
chunks = splitter.split_text("This is a long text that needs splitting into tokens.")

This splits text into chunks of 200 tokens without any overlap.

LangChain

from langchain.text_splitter import TokenTextSplitter

splitter = TokenTextSplitter(chunk_size=200, chunk_overlap=0)
chunks = splitter.split_text(text)

Sample Program

This example splits a short text into chunks of 10 tokens each, with 3 tokens overlapping. It prints each chunk so you can see how the text is divided.

LangChain

from langchain.text_splitter import TokenTextSplitter

text = "Langchain helps you build applications with language models. Token-based splitting breaks text into token chunks for better processing."

splitter = TokenTextSplitter(chunk_size=10, chunk_overlap=3)
chunks = splitter.split_text(text)

for i, chunk in enumerate(chunks, 1):
    print(f"Chunk {i}: {chunk}")

Important Notes

Token splitting depends on the tokenizer used by the language model.

Overlap helps keep context but increases total tokens processed.

Adjust chunk size and overlap based on your model's token limits and needs.

Summary

Token-based splitting breaks text into token-sized chunks for easier processing.

Use chunk size and overlap to control chunk length and context.

This method is useful when working with language models that have token limits.

Practice

1. What is the main purpose of token-based splitting in langchain?

easy

A. To translate text into different languages automatically

B. To combine multiple texts into one large string

C. To encrypt text data for security

D. To break text into smaller token-sized chunks for easier processing

Solution

1. Step 1: Understand token-based splitting

   Token-based splitting divides text into smaller pieces based on tokens, which are units of text like words or subwords.

2. Step 2: Identify the purpose

   This helps manage large texts by breaking them into manageable chunks for language models with token limits.

3. Final Answer:

   To break text into smaller token-sized chunks for easier processing -> Option D

4. Quick Check:

   Token-based splitting = chunking text [OK]

Hint: Token splitting means breaking text into small pieces [OK]

Common Mistakes:

• Confusing splitting with combining texts
• Thinking it encrypts or translates text
• Assuming it changes text meaning

2. Which of the following is the correct way to create a token-based text splitter in langchain?

easy

A. TextSplitter(chunk_size=100, overlap=20)

B. TokenTextSplitter(chunk_size=100, chunk_overlap=20)

C. TokenSplitter(size=100, overlap=20)

D. TokenTextSplitter(size=100, overlap=20)

Solution

1. Step 1: Recall the correct class name and parameters

   The correct class is TokenTextSplitter with parameters chunk_size and chunk_overlap.

2. Step 2: Match options with correct syntax

   TokenTextSplitter(chunk_size=100, chunk_overlap=20) uses TokenTextSplitter with chunk_size and chunk_overlap, which is correct.

3. Final Answer:

   TokenTextSplitter(chunk_size=100, chunk_overlap=20) -> Option B

4. Quick Check:

   Correct class and parameters = TokenTextSplitter(chunk_size=100, chunk_overlap=20) [OK]

Hint: Look for TokenTextSplitter with chunk_size and chunk_overlap [OK]

Common Mistakes:

• Using wrong class names like TokenSplitter
• Using incorrect parameter names like size or overlap
• Mixing parameter names or omitting chunk_ prefix

3. Given the following code snippet, what will be the length of the first chunk if the text has 250 tokens, chunk_size=100, and chunk_overlap=20?

splitter = TokenTextSplitter(chunk_size=100, chunk_overlap=20)
chunks = splitter.split_text(text)

medium

A. 100 tokens

B. 120 tokens

C. 80 tokens

D. 250 tokens

Solution

1. Step 1: Understand chunk size and overlap

   The chunk_size defines the max tokens per chunk, here 100 tokens.

2. Step 2: Determine first chunk length

   The first chunk starts at token 0 and includes 100 tokens; overlap applies to next chunks, not the first.

3. Final Answer:

   100 tokens -> Option A

4. Quick Check:

   First chunk = chunk_size = 100 tokens [OK]

Hint: First chunk length equals chunk_size, overlap affects next chunks [OK]

Common Mistakes:

• Adding overlap to first chunk size
• Assuming first chunk covers entire text
• Confusing overlap as extra tokens in first chunk

4. What is wrong with this code snippet for token-based splitting?

splitter = TokenTextSplitter(chunk_size=50, chunk_overlap=60)
chunks = splitter.split_text(text)

medium

A. split_text method does not exist

B. chunk_size must be a string, not a number

C. chunk_overlap cannot be larger than chunk_size

D. chunk_overlap must be zero

Solution

1. Step 1: Check parameter constraints

   chunk_overlap should not exceed chunk_size because overlap means shared tokens between chunks.

2. Step 2: Identify error in values

   Here, chunk_overlap=60 is greater than chunk_size=50, which is invalid and causes an error.

3. Final Answer:

   chunk_overlap cannot be larger than chunk_size -> Option C

4. Quick Check:

   Overlap ≤ chunk size rule violated [OK]

Hint: Overlap must be less or equal to chunk size [OK]

Common Mistakes:

• Thinking chunk_size must be string
• Assuming split_text method is missing
• Believing overlap must be zero always

5. You have a long document of 1000 tokens and want to split it into chunks of 200 tokens with 50 tokens overlap. How many chunks will you get using TokenTextSplitter(chunk_size=200, chunk_overlap=50)?

hard

A. 7 chunks

B. 6 chunks

C. 8 chunks

D. 5 chunks

Solution

1. Step 1: Calculate effective step size

   Each new chunk starts 200 - 50 = 150 tokens after the previous chunk start.

2. Step 2: Calculate number of chunks

   Number of chunks = ceil((1000 - 50) / 150) = ceil(950 / 150) = ceil(6.33) = 7 chunks.

3. Final Answer:

   7 chunks -> Option A

4. Quick Check:

   Chunks = ceil((total - overlap) / (chunk_size - overlap)) = 7 [OK]

Hint: Use step = chunk_size - overlap, then divide total tokens [OK]

Common Mistakes:

• Ignoring overlap in calculation
• Dividing total tokens by chunk_size directly
• Rounding down instead of up