Blog 2 — Fixed-Size Chunking: Fast but Blind to Meaning

Fixed-size chunking is the simplest way to divide a document.

The splitter receives a size limit, such as:

500 characters

or:

200 tokens

It then moves through the document and creates chunks around that size.

Document
   ↓
First 500 characters  → Chunk 1
Next 500 characters   → Chunk 2
Next 500 characters   → Chunk 3

A character is a letter, number, punctuation mark, or space. A token is a unit used by a language model and may contain one word or only part of a word.

Strengths of fixed-size chunking

Fixed-size chunking requires very little processing.

It is usually:

• fast
• easy to implement
• predictable
• easy to debug

It is also deterministic, which means the same document and settings produce the same chunks every time.

These properties make fixed-size chunking a useful baseline. A baseline is a simple starting solution used to compare later improvements.

The boundary problem

Fixed-size chunking mainly pays attention to length. In its simplest form, it does not understand sentences, topics, or document structure.

Consider the following text:

My name is Andy.
I work as a data engineer.
My favourite food is beef noodle soup.

A fixed boundary may produce:

Chunk 1:
My name is Andy.
I work as a data engineer.
My favourite food is

Chunk 2:
beef noodle soup.

The phrase “beef noodle soup” remains complete, but its relationship with Andy has been separated.

The situation becomes worse when a hard character boundary cuts through a word:

Chunk 1:
My name is An

Chunk 2:
dy. I work as a data engineer.

Not every fixed-size implementation makes such a hard cut. Some libraries use separators or tokenizers to reduce broken words. However, a size limit alone cannot guarantee that a complete idea remains in one chunk.

Adding overlap

Overlap is often combined with fixed-size chunking.

For example, a 500-token chunk might repeat the final 50 tokens at the beginning of the next chunk.

Chunk 1: tokens 0–500
Chunk 2: tokens 450–950
Chunk 3: tokens 900–1400

This gives information near the boundary another chance to appear in a complete chunk.

Overlap reduces some boundary problems, but it does not remove them. An important explanation may still be longer than the overlapping area.

Large overlap also creates disadvantages:

More overlap
→ more duplicated text
→ more embeddings
→ more storage
→ more repeated search results

Suitable use cases

Fixed-size chunking can work well when:

• the documents have a uniform format
• each record already contains complete information
• indexing speed is important
• a simple comparison baseline is needed

It is less suitable when sentences, paragraphs, or sections must remain together.

Design decision

Compared with treating a whole document as one retrieval unit, fixed-size chunking improves speed, predictability, and retrieval focus.

Its main weakness is meaning preservation.

Fixed-size chunking

Main priority:
speed and simplicity

Main risk:
arbitrary boundaries

The next strategy keeps the size limit but changes how boundaries are selected. Instead of cutting immediately when the limit is reached, recursive chunking first looks for a more natural place to split.

References

1. LangChain, Text Splitter Integrations: length-based chunking can use character counts or token counts to control chunk size. (Docs by LangChain)
2. LangChain, Splitting by Character: character-based splitting uses a chosen separator and measures chunk length by character count. (Docs by LangChain)
3. freeCodeCamp.org, Production RAG with LangChain & Vector Databases – Full Course. Watch the YouTube course. (YouTube)