SECAI Core · phase 7 of 15 · noun
Token limit
The maximum number of tokens an AI model can process in a single request
The Explain card
- Plain English
- A token limit is the maximum number of tokens a model can process in a single request, covering the prompt, retrieved context and the generated response.
- Example
- A RAG pipeline stuffs twenty retrieved documents into the prompt. The request exceeds the limit, so the system silently truncates the beginning, which happened to contain the safety instructions in the system prompt.
- Why it matters
- Limits shape what the model sees and what gets dropped. Attackers can pad inputs to push guardrails out of the window or exhaust quotas. Defenders manage context carefully and watch for truncation.
- Hook
- The token limit is the window size; whatever falls outside it never existed to the model.
Word knowledge
How the word is built, where it came from, and what it sits beside in memory.
In a sentence
The AI gateway counts subword units in the system prompt and the retrieved passages, then truncates the prompt to fit the model's token limit.
Why these words
- token Old English tācen, sign, evidence, mark supplies the unit being capped, so the bound is counted in model pieces of text rather than in characters, files, or requests
- limit Latin limes, limitis, boundary or frontier names the bound the phrase draws, turning a general cap-word into the head of a resource collocation
Where it came from
- Origin
- English computing collocation from Old English tācen (sign, evidence) and Latin limes, limitis (boundary, frontier)
- Entered the language
- late 2010s
- What changed
- Token had meant a sign or a piece of evidence, then a metal stand-in for a fare, then the smallest unit a compiler's lexer emits. Limit had meant a Roman frontier, then any bound on size or number. Transformer labs of the late 2010s joined the two words when models began advertising a hard count of subword units, and API references made that count a named knob on each request.
How it is spelled
- Pattern
- o before one consonant plus a vowel often spells long /oʊ/, as in token, open, and frozen
- Pattern
- each i in limit sits in a closed syllable and stays short, as in civic and finish
- Pattern
- the noun phrase stays two words as an open compound
- Breaks the pattern
- tokenize keeps a single k; English does not double it
- Breaks the pattern
- a hyphen appears only when the pair modifies another noun, as in token-limit check
Spelled like
- open
- frozen
- civic
Broken into chunks
-
token
- tokenize
- tokenizer
- betoken
-
lim
- unlimited
- delimit
- limitation
What it sits beside
Same subject
Same shape
- rate limit
- modality limit
- time limit
- speed limit
Gateway caps
Limit compounds
- rate limit
- time limit
- speed limit
Token family
- tokenizer
- tokenization
- context window
Where it sits in the deck
Phase 7: AI-Specific Runtime Controls: Guardrails, Limits, and Enforcement
Responsible AI principles must be enforced mechanically — learn the runtime controls that constrain model inputs, outputs, and resource consumption.