Skip to content

TokenBudgetPolicy

autogen.beta.policies.token_budget.TokenBudgetPolicy #

TokenBudgetPolicy(max_tokens, chars_per_token=4, transparent=False)

Keep events within a token budget.

Estimates tokens by character count. Retains most recent events first.

Source code in autogen/beta/policies/token_budget.py 
19
20
21
def __init__(self, max_tokens: int, chars_per_token: int = 4, transparent: bool = False) -> None:
    self._max_chars = max_tokens * chars_per_token
    self._transparent = transparent

name class-attribute instance-attribute #

name = 'token_budget'

apply async #

apply(prompts, events, context)
Source code in autogen/beta/policies/token_budget.py 
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
async def apply(
    self,
    prompts: list[str],
    events: list[BaseEvent],
    context: Context,
) -> tuple[list[str], list[BaseEvent]]:
    total_chars = sum(len(str(e)) for e in events)
    if total_chars <= self._max_chars:
        return prompts, events

    # Retain from the end, fitting within budget
    retained: list[BaseEvent] = []
    budget = self._max_chars
    for event in reversed(events):
        cost = len(str(event))
        if budget - cost < 0 and retained:
            break
        retained.append(event)
        budget -= cost
    retained.reverse()
    # Skip leading ToolResultsEvents whose matching tool_use was trimmed away.
    while retained and isinstance(retained[0], ToolResultsEvent):
        retained = retained[1:]

    if self._transparent:
        prompts = prompts + [f"[{self.name}] Showing {len(retained)} of {len(events)} events (token budget)."]
    return prompts, retained