The AI Handbook Open in the app →

Tokenizer

Inference & Prompting · Intermediate · 4 min read

What is it?

A tokenizer is the tool that splits text into tokens — the small chunks (word pieces) a language model actually reads and generates.

Explain like I'm 5

A model can't read letters directly, so the tokenizer chops text into bite-size pieces it understands, like cutting food before a toddler eats it.

Why was it created?

Models work with numbers, not raw text, and whole words are too many to list. Tokenizers were created to break text into a manageable, reusable vocabulary of sub-word pieces.

Where is it used?

  • Every LLM prompt and response
  • Counting cost and context usage
  • Preparing training data
  • Multilingual and code handling

Why should developers care?

Tokens determine cost, context limits, and quirks like why models miscount letters — so understanding tokenization explains a lot of real LLM behavior.

How does it work?

The tokenizer maps text to token IDs using a learned vocabulary of sub-word units (common methods: BPE, WordPiece). Frequent words become single tokens; rare or novel words split into pieces. The model reads these IDs and outputs token IDs, which are decoded back to text.

Real-world example

'unbelievable' might become 'un', 'believ', 'able' — three tokens — while 'cat' is one; this is why your bill and context usage depend on tokens, not words.

Common use cases

  • Estimating token cost
  • Fitting text within a context window
  • Chunking documents for retrieval
  • Debugging odd model behavior

Advantages

  • Compact, reusable vocabulary
  • Handles unknown words via sub-words
  • Works across languages and code
  • Enables numeric input for models

Disadvantages

  • Token boundaries can be unintuitive
  • Some languages use far more tokens
  • Causes character-level blind spots
  • Different models tokenize differently

When should you use it?

Whenever you need to reason about cost, context limits, or how a model chunks and reads your text.

When should you avoid it?

You rarely avoid it — but you don't need to hand-tune tokenization for typical app development.

Alternatives

Character-level encodingWord-level tokenizationByte-level encoding

Related terms

TokenContext WindowLarge Language ModelEmbeddings

Interview questions

Beginner

  • What does a tokenizer produce from text?
  • Why isn't one token always one word?

Intermediate

  • Why can an LLM struggle to count letters in a word?
  • How does sub-word tokenization handle unseen words?

Senior

  • How does byte-pair encoding build its vocabulary?
  • Why can the same text cost more tokens in one language than another?

Common misconceptions

  • "One token equals one word" — tokens are often word fragments.
  • "Tokenization doesn't affect quality" — it shapes cost, limits, and some failure modes.

Fun facts

  • A rough rule of thumb: ~4 characters per token in English.
  • Poor tokenization is a big reason models fumble tasks like counting characters.

Timeline

  • 2016 — Byte-pair encoding popularized for neural translation
  • 2020s — Sub-word tokenizers standard across LLMs

Learning resources

Quick summary

A tokenizer splits text into sub-word tokens the model reads and writes, which drives cost, context limits, and some model quirks.

Cheat sheet

  • Splits text into tokens (word pieces)
  • ~4 chars/token in English
  • Drives cost & context usage
  • Explains character-counting errors

If you remember only one thing

A tokenizer chops text into the sub-word pieces a model actually reads — which is why cost and context are measured in tokens, not words.

Related tools