Compare commits

...

1 Commits

Author SHA1 Message Date
Devin AI
3bfa1c6559 Fix issue #3454: Add proactive context length checking to prevent empty LLM responses
- Add _check_context_length_before_call() method to CrewAgentExecutor
- Proactively check estimated token count before LLM calls in _invoke_loop
- Use character-based estimation (chars / 4) to approximate token count
- Call existing _handle_context_length() when context window would be exceeded
- Add comprehensive tests covering proactive handling and token estimation
- Prevents empty responses from providers like DeepInfra that don't throw exceptions

Co-Authored-By: João <joao@crewai.com>
2025-09-05 16:05:35 +00:00
2 changed files with 90 additions and 0 deletions

View File

@@ -112,6 +112,8 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
try:
while not isinstance(formatted_answer, AgentFinish):
if not self.request_within_rpm_limit or self.request_within_rpm_limit():
self._check_context_length_before_call()
answer = self.llm.call(
self.messages,
callbacks=self.callbacks,
@@ -327,6 +329,19 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
)
]
def _check_context_length_before_call(self) -> None:
total_chars = sum(len(msg.get("content", "")) for msg in self.messages)
estimated_tokens = total_chars // 4
context_window_size = self.llm.get_context_window_size()
if estimated_tokens > context_window_size:
self._printer.print(
content=f"Estimated token count ({estimated_tokens}) exceeds context window ({context_window_size}). Handling proactively.",
color="yellow",
)
self._handle_context_length()
def _handle_context_length(self) -> None:
if self.respect_context_window:
self._printer.print(

View File

@@ -1625,3 +1625,78 @@ def test_agent_with_knowledge_sources():
# Assert that the agent provides the correct information
assert "red" in result.raw.lower()
def test_proactive_context_length_handling_prevents_empty_response():
"""Test that proactive context length checking prevents empty LLM responses."""
agent = Agent(
role="test role",
goal="test goal",
backstory="test backstory",
sliding_context_window=True,
)
long_input = "This is a very long input that should exceed the context window. " * 1000
with patch.object(agent.llm, 'get_context_window_size', return_value=100):
with patch.object(agent.agent_executor, '_handle_context_length') as mock_handle:
with patch.object(agent.llm, 'call', return_value="Proper response after summarization"):
agent.agent_executor.messages = [
{"role": "user", "content": long_input}
]
task = Task(
description="Process this long input",
expected_output="A response",
agent=agent,
)
result = agent.execute_task(task)
mock_handle.assert_called()
assert result and result.strip() != ""
def test_proactive_context_length_handling_with_no_summarization():
"""Test proactive context length checking when summarization is disabled."""
agent = Agent(
role="test role",
goal="test goal",
backstory="test backstory",
sliding_context_window=False,
)
long_input = "This is a very long input. " * 1000
with patch.object(agent.llm, 'get_context_window_size', return_value=100):
agent.agent_executor.messages = [
{"role": "user", "content": long_input}
]
with pytest.raises(SystemExit):
agent.agent_executor._check_context_length_before_call()
def test_context_length_estimation():
"""Test the token estimation logic."""
agent = Agent(
role="test role",
goal="test goal",
backstory="test backstory",
)
agent.agent_executor.messages = [
{"role": "user", "content": "Short message"},
{"role": "assistant", "content": "Another short message"},
]
with patch.object(agent.llm, 'get_context_window_size', return_value=10):
with patch.object(agent.agent_executor, '_handle_context_length') as mock_handle:
agent.agent_executor._check_context_length_before_call()
mock_handle.assert_not_called()
with patch.object(agent.llm, 'get_context_window_size', return_value=5):
with patch.object(agent.agent_executor, '_handle_context_length') as mock_handle:
agent.agent_executor._check_context_length_before_call()
mock_handle.assert_called()