Skip to content

The agent engineering platform.

core: Improve OutputParser error messaging when model output is truncated (max_tokens) (#30936)

Ahmed Tammaade56c31672fa
diff --git a/libs/core/langchain_core/output_parsers/openai_tools.py b/libs/core/langchain_core/output_parsers/openai_tools.py
index 254d11eca1..b4f845bdca 100644
--- a/libs/core/langchain_core/output_parsers/openai_tools.py
+++ b/libs/core/langchain_core/output_parsers/openai_tools.py
@@ -2,6 +2,7 @@
import copy
import json
+import logging
from json import JSONDecodeError
from typing import Annotated, Any, Optional
@@ -16,6 +17,8 @@ from langchain_core.outputs import ChatGeneration, Generation
from langchain_core.utils.json import parse_partial_json
from langchain_core.utils.pydantic import TypeBaseModel
+logger = logging.getLogger(__name__)
+
def parse_tool_call(
raw_tool_call: dict[str, Any],
@@ -250,6 +253,14 @@ class JsonOutputKeyToolsParser(JsonOutputToolsParser):
return parsed_result
+# Common cause of ValidationError is truncated output due to max_tokens.
+_MAX_TOKENS_ERROR = (
+ "Output parser received a `max_tokens` stop reason. "
+ "The output is likely incomplete—please increase `max_tokens` "
+ "or shorten your prompt."
+)
+
+
class PydanticToolsParser(JsonOutputToolsParser):
"""Parse tools from OpenAI response."""
@@ -296,6 +307,14 @@ class PydanticToolsParser(JsonOutputToolsParser):
except (ValidationError, ValueError):
if partial:
continue
+ has_max_tokens_stop_reason = any(
+ generation.message.response_metadata.get("stop_reason")
+ == "max_tokens"
+ for generation in result
+ if isinstance(generation, ChatGeneration)
+ )
+ if has_max_tokens_stop_reason:
+ logger.exception(_MAX_TOKENS_ERROR)
raise
if self.first_tool_only:
return pydantic_objects[0] if pydantic_objects else None
diff --git a/libs/core/tests/unit_tests/output_parsers/test_openai_tools.py b/libs/core/tests/unit_tests/output_parsers/test_openai_tools.py
index c63dd693d7..992e5c48c5 100644
--- a/libs/core/tests/unit_tests/output_parsers/test_openai_tools.py
+++ b/libs/core/tests/unit_tests/output_parsers/test_openai_tools.py
@@ -2,7 +2,7 @@ from collections.abc import AsyncIterator, Iterator
from typing import Any
import pytest
-from pydantic import BaseModel, Field
+from pydantic import BaseModel, Field, ValidationError
from langchain_core.messages import (
AIMessage,
@@ -635,3 +635,24 @@ def test_parse_with_different_pydantic_1_proper() -> None:
forecast="Sunny",
)
]
+
+
+def test_max_tokens_error(caplog: Any) -> None:
+ parser = PydanticToolsParser(tools=[NameCollector], first_tool_only=True)
+ input = AIMessage(
+ content="",
+ tool_calls=[
+ {
+ "id": "call_OwL7f5PE",
+ "name": "NameCollector",
+ "args": {"names": ["suz", "jerm"]},
+ }
+ ],
+ response_metadata={"stop_reason": "max_tokens"},
+ )
+ with pytest.raises(ValidationError):
+ _ = parser.invoke(input)
+ assert any(
+ "`max_tokens` stop reason" in msg and record.levelname == "ERROR"
+ for record, msg in zip(caplog.records, caplog.messages)
+ )