Skip to content

The agent engineering platform.

fix(core): adjust cap when scaling approximate token counts (#35017)

ccurme032d01dd0ff1
diff --git a/libs/core/langchain_core/messages/utils.py b/libs/core/langchain_core/messages/utils.py
index 0d1c673901..4d89fc10a3 100644
--- a/libs/core/langchain_core/messages/utils.py
+++ b/libs/core/langchain_core/messages/utils.py
@@ -2320,7 +2320,7 @@ def count_tokens_approximately(
and approx_at_last_ai > 0
):
scale_factor = last_ai_total_tokens / approx_at_last_ai
- token_count *= min(1.5, max(1.0, scale_factor))
+ token_count *= min(1.25, max(1.0, scale_factor))
# round up once more time in case extra_tokens_per_message is a float
return math.ceil(token_count)
diff --git a/libs/core/tests/unit_tests/messages/test_utils.py b/libs/core/tests/unit_tests/messages/test_utils.py
index ae52893030..cdeb7d764c 100644
--- a/libs/core/tests/unit_tests/messages/test_utils.py
+++ b/libs/core/tests/unit_tests/messages/test_utils.py
@@ -1615,7 +1615,7 @@ def test_count_tokens_approximately_usage_metadata_scaling() -> None:
scaled = count_tokens_approximately(messages, use_usage_metadata_scaling=True)
ratio = scaled / unscaled
- assert 1 <= ratio <= 1.5
+ assert 1 <= round(ratio, 1) <= 1.2 # we ceil scale token counts, so can be > 1.2
messages.extend([ToolMessage("text", tool_call_id="abc123")] * 3)
@@ -1631,7 +1631,7 @@ def test_count_tokens_approximately_usage_metadata_scaling() -> None:
# And the scaled total should be the unscaled total multiplied by the same ratio.
# ratio = 200 / unscaled (as of last AI message)
expected_scaled_extended = math.ceil(unscaled_extended * ratio)
- assert scaled_extended == expected_scaled_extended
+ assert scaled_extended <= expected_scaled_extended <= scaled_extended + 1
def test_count_tokens_approximately_usage_metadata_scaling_model_provider() -> None: