This removes any specific model from the "Setting up your LLM" guide,
but provides examples for the top-3 providers.
This section also conflated "model selection" with "model
configuration", where configuration is provider-specific, so I've
focused this first section on just model selection, deferring the config
to the "provider" section that follows.
Co-authored-by: Tony Kipkemboi <iamtonykipkemboi@gmail.com>
This commit adds a new crew field called parent_flow, evaluated when the Crew
instance is instantiated. The stacktrace is traversed to look up if the caller
is an instance of Flow, and if so, it fills in the field.
Other alternatives were considered, such as a global context or even a new
field to be manually filled, however, this is the most magical solution that
was thread-safe and did not require public API changes.
* fix: support to reset memories after changing Crew's embedder
The sources must not be added while initializing the Knowledge otherwise we could not reset it
* chore: improve reset memory feedback
Previously, even when no memories were actually erased, we logged that they had been. From now on, the log will specify which memory has been reset.
* feat: improve get_crew discovery from a single file
Crew instances can now be discovered from any function or method with a return type annotation of -> Crew, as well as from module-level attributes assigned to a Crew instance. Additionally, crews can be retrieved from within a Flow
* refactor: make add_sources a public method from Knowledge
* build(dev): add pytest-randomly dependency
By randomizing the test execution order, this helps identify tests
that unintentionally depend on shared state or specific execution
order, which can lead to flaky or unreliable test behavior.
* build(dev): add pytest-timeout
This will prevent a test from running indefinitely
* test: block external requests in CI and set default 10s timeout per test
* test: adding missing cassettes
We notice that those cassettes are missing after enabling block-network on CI
* test: increase tests timeout on CI
* test: fix flaky test ValueError: Circular reference detected (id repeated)
* fix: prevent crash when event handler raises exception
Previously, if a registered event handler raised an exception during execution,
it could crash the entire application or interrupt the event dispatch process.
This change wraps handler execution in a try/except block within the `emit` method,
ensuring that exceptions are caught and logged without affecting other handlers or flow.
This improves the resilience of the event bus, especially when handling third-party
or temporary listeners.
* feat: support to define a guardrail task no-code
* feat: add auto-discovery for Guardrail code execution mode
* feat: handle malformed or invalid response from CodeInterpreterTool
* feat: allow to set unsafe_mode from Guardrail task
* feat: renaming GuardrailTask to TaskGuardrail
* feat: ensure guardrail is callable while initializing Task
* feat: remove Docker availability check from TaskGuardrail
The CodeInterpreterTool already ensures compliance with this requirement.
* refactor: replace if/raise with assert
For this use case `assert` is more appropriate choice
* test: remove useless or duplicated test
* fix: attempt to fix type-checker
* feat: support to define a task guardrail using YAML config
* refactor: simplify TaskGuardrail to use LLM for validation, no code generation
* docs: update TaskGuardrail doc strings
* refactor: drop task paramenter from TaskGuardrail
This parameter was used to get the model from the `task.agent` which is a quite bit redudant since we could propagate the llm directly
Add `__init__.py` files to 20 directories to conform with Python package standards. This ensures directories are properly recognized as packages, enabling cleaner imports.
This commit fixes a bug where changing logging level would be overriden
by `src/crewai/project/crew_base.py`. For example, the following snippet
on top of a crew or flow would not work:
```python
logging.basicConfig(
level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)
```
Crews and flows should be able to set their own log level, without being
overriden by CrewAI library code.
* Fix issue #2402: Handle missing templates gracefully
Co-Authored-By: Joe Moura <joao@crewai.com>
* Fix import sorting in test files
Co-Authored-By: Joe Moura <joao@crewai.com>
* Bluit in top of devin-ai integration
* Fixed test cases
* Fixed test cases
* fixed linting issue
* Added docs
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Joe Moura <joao@crewai.com>
* build(litellm): upgrade LiteLLM to latest version
* fix: update filtered logs from LiteLLM
* Fix for a missing backtick
---------
Co-authored-by: Mike Plachta <mike@crewai.com>
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
* added gpt4.1 models and gemini 2.0 and 2.5 models
* added flash model
* Updated test fun to all models
* Added Gemma3 test cases and passed all google test case
* added gemini 2.5 flash
* added gpt4.1 models and gemini 2.0 and 2.5 models
* added flash model
* Updated test fun to all models
* Added Gemma3 test cases and passed all google test case
* added gemini 2.5 flash
* added gpt4.1 models and gemini 2.0 and 2.5 models
* added flash model
* Updated test fun to all models
* Added Gemma3 test cases and passed all google test case
* added gemini 2.5 flash
* test: add missing cassettes
* test: ignore authorization key from gemini/gemma3 request
---------
Co-authored-by: Lucas Gomide <lucaslg200@gmail.com>
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
* Enhance knowledge management in CrewAI
- Added `KnowledgeConfig` class to configure knowledge retrieval parameters such as `limit` and `score_threshold`.
- Updated `Agent` and `Crew` classes to utilize the new knowledge configuration for querying knowledge sources.
- Enhanced documentation to clarify the addition of knowledge sources at both agent and crew levels.
- Introduced new tips in documentation to guide users on knowledge source management and configuration.
* Refactor knowledge configuration parameters in CrewAI
- Renamed `limit` to `results_limit` in `KnowledgeConfig`, `query_knowledge`, and `query` methods for consistency and clarity.
- Updated related documentation to reflect the new parameter name, ensuring users understand the configuration options for knowledge retrieval.
* Refactor agent tests to utilize mock knowledge storage
- Updated test cases in `agent_test.py` to use `KnowledgeStorage` for mocking knowledge sources, enhancing test reliability and clarity.
- Renamed `limit` to `results_limit` in `KnowledgeConfig` for consistency with recent changes.
- Ensured that knowledge queries are properly mocked to return expected results during tests.
* Add VCR support for agent tests with query limits and score thresholds
- Introduced `@pytest.mark.vcr` decorator in `agent_test.py` for tests involving knowledge sources, ensuring consistent recording of HTTP interactions.
- Added new YAML cassette files for `test_agent_with_knowledge_sources_with_query_limit_and_score_threshold` and `test_agent_with_knowledge_sources_with_query_limit_and_score_threshold_default`, capturing the expected API responses for these tests.
- Enhanced test reliability by utilizing VCR to manage external API calls during testing.
* Update documentation to format parameter names in code style
- Changed the formatting of `results_limit` and `score_threshold` in the documentation to use code style for better clarity and emphasis.
- Ensured consistency in documentation presentation to enhance user understanding of configuration options.
* Enhance KnowledgeConfig with field descriptions
- Updated `results_limit` and `score_threshold` in `KnowledgeConfig` to use Pydantic's `Field` for improved documentation and clarity.
- Added descriptions to both parameters to provide better context for their usage in knowledge retrieval configuration.
* docstrings added
* feat: add OpenAI agent adapter implementation
- Introduced OpenAIAgentAdapter class to facilitate interaction with OpenAI Assistants.
- Implemented methods for task execution, tool configuration, and response processing.
- Added support for converting CrewAI tools to OpenAI format and handling delegation tools.
* created an adapter for the delegate and ask_question tools
* delegate and ask_questions work and it delegates to crewai agents*
* refactor: introduce OpenAIAgentToolAdapter for tool management
- Created OpenAIAgentToolAdapter class to encapsulate tool configuration and conversion for OpenAI Assistant.
- Removed tool configuration logic from OpenAIAgentAdapter and integrated it into the new adapter.
- Enhanced the tool conversion process to ensure compatibility with OpenAI's requirements.
* feat: implement BaseAgentAdapter for agent integration
- Introduced BaseAgentAdapter as an abstract base class for agent adapters in CrewAI.
- Defined common interface and methods for configuring tools and structured output.
- Updated OpenAIAgentAdapter to inherit from BaseAgentAdapter, enhancing its structure and functionality.
* feat: add LangGraph agent and tool adapter for CrewAI integration
- Introduced LangGraphAgentAdapter to facilitate interaction with LangGraph agents.
- Implemented methods for task execution, context handling, and tool configuration.
- Created LangGraphToolAdapter to convert CrewAI tools into LangGraph-compatible format.
- Enhanced error handling and logging for task execution and streaming processes.
* feat: enhance LangGraphToolAdapter and improve conversion instructions
- Added type hints for better clarity and type checking in LangGraphToolAdapter.
- Updated conversion instructions to ensure compatibility with optional LLM checks.
* feat: integrate structured output handling in LangGraph and OpenAI agents
- Added LangGraphConverterAdapter for managing structured output in LangGraph agents.
- Enhanced LangGraphAgentAdapter to utilize the new converter for system prompt and task execution.
- Updated LangGraphToolAdapter to use StructuredTool for better compatibility.
- Introduced OpenAIConverterAdapter for structured output management in OpenAI agents.
- Improved task execution flow in OpenAIAgentAdapter to incorporate structured output configuration and post-processing.
* feat: implement BaseToolAdapter for tool integration
- Introduced BaseToolAdapter as an abstract base class for tool adapters in CrewAI.
- Updated LangGraphToolAdapter and OpenAIAgentToolAdapter to inherit from BaseToolAdapter, enhancing their structure and functionality.
- Improved tool configuration methods to support better integration with various frameworks.
- Added type hints and documentation for clarity and maintainability.
* feat: enhance OpenAIAgentAdapter with configurable agent properties
- Refactored OpenAIAgentAdapter to accept agent configuration as an argument.
- Introduced a method to build a system prompt for the OpenAI agent, improving task execution context.
- Updated initialization to utilize role, goal, and backstory from kwargs, enhancing flexibility in agent setup.
- Improved tool handling and integration within the adapter.
* feat: enhance agent adapters with structured output support
- Introduced BaseConverterAdapter as an abstract class for structured output handling.
- Implemented LangGraphConverterAdapter and OpenAIConverterAdapter to manage structured output in their respective agents.
- Updated BaseAgentAdapter to accept an agent configuration dictionary during initialization.
- Enhanced LangGraphAgentAdapter to utilize the new converter and improved tool handling.
- Added methods for configuring structured output and enhancing system prompts in converter adapters.
* refactor: remove _parse_tools method from OpenAIAgentAdapter and BaseAgent
- Eliminated the _parse_tools method from OpenAIAgentAdapter and its abstract declaration in BaseAgent.
- Cleaned up related test code in MockAgent to reflect the removal of the method.
* also removed _parse_tools here as not used
* feat: add dynamic import handling for LangGraph dependencies
- Implemented conditional imports for LangGraph components to handle ImportError gracefully.
- Updated LangGraphAgentAdapter initialization to check for LangGraph availability and raise an informative error if dependencies are missing.
- Enhanced the agent adapter's robustness by ensuring it only initializes components when the required libraries are present.
* fix: improve error handling for agent adapters
- Updated LangGraphAgentAdapter to raise an ImportError with a clear message if LangGraph dependencies are not installed.
- Refactored OpenAIAgentAdapter to include a similar check for OpenAI dependencies, ensuring robust initialization and user guidance for missing libraries.
- Enhanced overall error handling in agent adapters to prevent runtime issues when dependencies are unavailable.
* refactor: enhance tool handling in agent adapters
- Updated BaseToolAdapter to initialize original and converted tools in the constructor.
- Renamed method `all_tools` to `tools` for clarity in BaseToolAdapter.
- Added `sanitize_tool_name` method to ensure tool names are API compatible.
- Modified LangGraphAgentAdapter to utilize the updated tool handling and ensure proper tool configuration.
- Refactored LangGraphToolAdapter to streamline tool conversion and ensure consistent naming conventions.
* feat: emit AgentExecutionCompletedEvent in agent adapters
- Added emission of AgentExecutionCompletedEvent in both LangGraphAgentAdapter and OpenAIAgentAdapter to signal task completion.
- Enhanced event handling to include agent, task, and output details for better tracking of execution results.
* docs: Enhance BaseConverterAdapter documentation
- Added a detailed docstring to the BaseConverterAdapter class, outlining its purpose and the expected functionality for all converter adapters.
- Updated the post_process_result method's docstring to specify the expected format of the result as a string.
* docs: Add comprehensive guide for bringing custom agents into CrewAI
- Introduced a new documentation file detailing the process of integrating custom agents using the BaseAgentAdapter, BaseToolAdapter, and BaseConverter.
- Included step-by-step instructions for creating custom adapters, configuring tools, and handling structured output.
- Provided examples for implementing adapters for various frameworks, enhancing the usability of CrewAI for developers.
* feat: Introduce adapted_agent flag in BaseAgent and update BaseAgentAdapter initialization
- Added an `adapted_agent` boolean field to the BaseAgent class to indicate if the agent is adapted.
- Updated the BaseAgentAdapter's constructor to pass `adapted_agent=True` to the superclass, ensuring proper initialization of the new field.
* feat: Enhance LangGraphAgentAdapter to support optional agent configuration
- Updated LangGraphAgentAdapter to conditionally apply agent configuration when creating the agent graph, allowing for more flexible initialization.
- Modified LangGraphToolAdapter to ensure only instances of BaseTool are converted, improving tool compatibility and handling.
* feat: Introduce OpenAIConverterAdapter for structured output handling
- Added OpenAIConverterAdapter to manage structured output conversion for OpenAI agents, enhancing their ability to process and format results.
- Updated OpenAIAgentAdapter to utilize the new converter for configuring structured output and post-processing results.
- Removed the deprecated get_output_converter method from OpenAIAgentAdapter.
- Added unit tests for BaseAgentAdapter and BaseToolAdapter to ensure proper functionality and integration of new features.
* feat: Enhance tool adapters to support asynchronous execution
- Updated LangGraphToolAdapter and OpenAIAgentToolAdapter to handle asynchronous tool execution by checking if the output is awaitable.
- Introduced `inspect` import to facilitate the awaitability check.
- Refactored tool wrapper functions to ensure proper handling of both synchronous and asynchronous tool results.
* fix: Correct method definition syntax and enhance tool adapter implementation
- Updated the method definition for `configure_structured_output` to include the `def` keyword for clarity.
- Added an asynchronous tool wrapper to ensure tools can operate in both synchronous and asynchronous contexts.
- Modified the constructor of the custom converter adapter to directly assign the agent adapter, improving clarity and functionality.
* linted
* refactor: Improve tool processing logic in BaseAgent
- Added a check to return an empty list if no tools are provided.
- Simplified the tool attribute validation by using a list of required attributes.
- Removed commented-out abstract method definition for clarity.
* refactor: Simplify tool handling in agent adapters
- Changed default value of `tools` parameter in LangGraphAgentAdapter to None for better handling of empty tool lists.
- Updated tool initialization in both LangGraphAgentAdapter and OpenAIAgentAdapter to directly pass the `tools` parameter, removing unnecessary list handling.
- Cleaned up commented-out code in OpenAIConverterAdapter to improve readability.
* refactor: Remove unused stream_task method from LangGraphAgentAdapter
- Deleted the `stream_task` method from LangGraphAgentAdapter to streamline the code and eliminate unnecessary complexity.
- This change enhances maintainability by focusing on essential functionalities within the agent adapter.
* feat: unblock LLM(stream=True) to work with tools
* feat: replace pytest-vcr by pytest-recording
1. pytest-vcr does not support httpx - which LiteLLM uses for streaming responses.
2. pytest-vcr is no longer maintained, last commit 6 years ago :fist::skin-tone-4:
3. pytest-recording supports modern request libraries (including httpx) and actively maintained
* refactor: remove @skip_streaming_in_ci
Since we have fixed streaming response issue we can remove this @skip_streaming_in_ci
---------
Co-authored-by: Lorenze Jay <63378463+lorenzejay@users.noreply.github.com>
* Adjust checking for callable crew object.
Changes back to how it was being done before.
Fixes#2307
* Fix specific memory reset errors.
When not initiated, the function should raise
the "memory system is not initialized" RuntimeError.
* Remove print statement
* Fixes test case
---------
Co-authored-by: Carlos Souza <carloshrsouza@gmail.com>
* Fix#2551: Add Huggingface to provider list in CLI
Co-Authored-By: Joe Moura <joao@crewai.com>
* Update Huggingface API key name to HF_TOKEN and remove base URL prompt
Co-Authored-By: Joe Moura <joao@crewai.com>
* Update Huggingface API key name to HF_TOKEN in documentation
Co-Authored-By: Joe Moura <joao@crewai.com>
* Fix import sorting in test_constants.py
Co-Authored-By: Joe Moura <joao@crewai.com>
* Fix import order in test_constants.py
Co-Authored-By: Joe Moura <joao@crewai.com>
* Fix import formatting in test_constants.py
Co-Authored-By: Joe Moura <joao@crewai.com>
* Skip failing tests in Python 3.11 due to VCR cassette issues
Co-Authored-By: Joe Moura <joao@crewai.com>
* Fix import order in knowledge_test.py
Co-Authored-By: Joe Moura <joao@crewai.com>
* Revert skip decorators to check if tests are flaky
Co-Authored-By: Joe Moura <joao@crewai.com>
* Restore skip decorators for tests with VCR cassette issues in Python 3.11
Co-Authored-By: Joe Moura <joao@crewai.com>
* revert skip pytest decorators
* Remove import sys and skip decorators from test files
Co-Authored-By: Joe Moura <joao@crewai.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Joe Moura <joao@crewai.com>
Co-authored-by: Lucas Gomide <lucaslg200@gmail.com>
* feat: support defining any memory in an isolated way
This change makes it easier to use a specific memory type without unintentionally enabling all others.
Previously, setting memory=True would implicitly configure all available memories (like LTM and STM), which might not be ideal in all cases. For example, when building a chatbot that only needs an external memory, users were forced to also configure LTM and STM — which rely on default OpenAPI embeddings — even if they weren’t needed.
With this update, users can now define a single memory in isolation, making the configuration process simpler and more flexible.
* feat: add tests to ensure we are able to use contextual memory by set individual memories
* docs: enhance memory documentation
* feat: warn when long-term memory is defined but entity memory is not
* fix: Correctly copy memory objects during crew training (#2593)
Co-Authored-By: Joe Moura <joao@crewai.com>
* style: Fix import order in tests/crew_test.py
Co-Authored-By: Joe Moura <joao@crewai.com>
* fix: Rely on validator for memory copy, update test assertions
Removes manual deep copy of memory objects in Crew.copy().
The Pydantic model_validator 'create_crew_memory' handles the
initialization of new memory instances for the copied crew.
Updates test_crew_copy_with_memory assertions to verify that
the private memory attributes (_short_term_memory, etc.) are
correctly initialized as new instances in the copied crew.
Co-Authored-By: Joe Moura <joao@crewai.com>
* Revert "fix: Rely on validator for memory copy, update test assertions"
This reverts commit 8702bf1e34.
* fix: Re-add manual deep copy for all memory types in Crew.copy
Addresses feedback on PR #2594 to ensure all memory objects
(short_term, long_term, entity, external, user) are correctly
deep copied using model_copy(deep=True).
Also simplifies the test case to directly verify the copy behavior
instead of relying on the train method.
Co-Authored-By: Joe Moura <joao@crewai.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Joe Moura <joao@crewai.com>
* fix: use mem0_local_config instead of config in Memory.from_config (#2587)
Co-Authored-By: Joe Moura <joao@crewai.com>
* refactor: consolidate tests as per PR feedback
Co-Authored-By: Joe Moura <joao@crewai.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: Joe Moura <joao@crewai.com>
