Class: RubyLLM::Toolbox::Truncator

Inherits:
Object
  • Object
show all
Defined in:
lib/ruby_llm/toolbox/truncator.rb

Overview

Token-budgets a string so a single tool result can never blow up the context window. Keeps the head and the tail (the two ends an agent most often needs) and elides the middle. Counting goes through ruby_llm-tokenizer; if no tokenizer backend is available for the model, it falls back to a ~4-chars-per-token heuristic so a tool never crashes just because a tokenizer gem isn’t installed.

Constant Summary collapse

MARKER_RESERVE =

tokens held back for the elision marker

16

Instance Method Summary collapse

Constructor Details

#initialize(model:, max_tokens:) ⇒ Truncator

Returns a new instance of Truncator.



16
17
18
19
 
# File 'lib/ruby_llm/toolbox/truncator.rb', line 16

def initialize(model:, max_tokens:)
  @model = model
  @max_tokens = [max_tokens.to_i, 1].max
end

Instance Method Details

#call(text) ⇒ Object 



21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
 
# File 'lib/ruby_llm/toolbox/truncator.rb', line 21

def call(text)
  text = text.to_s
  total = count(text)
  return text if total <= @max_tokens

  budget      = [@max_tokens - MARKER_RESERVE, 2].max
  head_budget = (budget * 0.6).floor
  tail_budget = budget - head_budget

  head = take(text, head_budget, from: :head)
  tail = take(text, tail_budget, from: :tail)
  cut  = total - count(head) - count(tail)

  "#{head}\n\n…[truncated ~#{cut} tokens]…\n\n#{tail}"
end