Pipeline

What is a Pipeline?¶

A pipeline in crewAI represents a structured workflow that allows for the sequential or parallel execution of multiple crews. It provides a way to organize complex processes involving multiple stages, where the output of one stage can serve as input for subsequent stages.

Key Terminology¶

Understanding the following terms is crucial for working effectively with pipelines:

Stage: A distinct part of the pipeline, which can be either sequential (a single crew) or parallel (multiple crews executing concurrently).
Kickoff: A specific execution of the pipeline for a given set of inputs, representing a single instance of processing through the pipeline.
Branch: Parallel executions within a stage (e.g., concurrent crew operations).
Trace: The journey of an individual input through the entire pipeline, capturing the path and transformations it undergoes.

Example pipeline structure:

crew1 >> [crew2, crew3] >> crew4

This represents a pipeline with three stages:

A sequential stage (crew1)
A parallel stage with two branches (crew2 and crew3 executing concurrently)
Another sequential stage (crew4)

Each input creates its own kickoff, flowing through all stages of the pipeline. Multiple kickoffs can be processed concurrently, each following the defined pipeline structure.

Pipeline Attributes¶

Attribute	Parameters	Description
Stages	`stages`	A list of `PipelineStage` (crews, lists of crews, or routers) representing the stages to be executed in sequence.

Creating a Pipeline¶

When creating a pipeline, you define a series of stages, each consisting of either a single crew or a list of crews for parallel execution. The pipeline ensures that each stage is executed in order, with the output of one stage feeding into the next.

Example: Assembling a Pipeline¶

from crewai import Crew, Process, Pipeline

# Define your crews
research_crew = Crew(
    agents=[researcher],
    tasks=[research_task],
    process=Process.sequential
)

analysis_crew = Crew(
    agents=[analyst],
    tasks=[analysis_task],
    process=Process.sequential
)

writing_crew = Crew(
    agents=[writer],
    tasks=[writing_task],
    process=Process.sequential
)

# Assemble the pipeline
my_pipeline = Pipeline(
    stages=[research_crew, analysis_crew, writing_crew]
)

Pipeline Methods¶

Method	Description
kickoff	Executes the pipeline, processing all stages and returning the results. This method initiates one or more kickoffs through the pipeline, handling the flow of data between stages.
process_runs	Runs the pipeline for each input provided, handling the flow and transformation of data between stages.

Pipeline Output¶

Understanding Pipeline Outputs

The output of a pipeline in the crewAI framework is encapsulated within the PipelineKickoffResult class. This class provides a structured way to access the results of the pipeline's execution, including various formats such as raw strings, JSON, and Pydantic models.

Pipeline Output Attributes¶

Attribute	Parameters	Type	Description
ID	`id`	`UUID4`	A unique identifier for the pipeline output.
Run Results	`run_results`	`List[PipelineRunResult]`	A list of `PipelineRunResult` objects, each representing the output of a single run through the pipeline.

Pipeline Output Methods¶

Method/Property	Description
add_run_result	Adds a `PipelineRunResult` to the list of run results.

Pipeline Run Result Attributes¶

Attribute	Parameters	Type	Description
ID	`id`	`UUID4`	A unique identifier for the run result.
Raw	`raw`	`str`	The raw output of the final stage in the pipeline kickoff.
Pydantic	`pydantic`	`Any`	A Pydantic model object representing the structured output of the final stage, if applicable.
JSON Dict	`json_dict`	`Union[Dict[str, Any], None]`	A dictionary representing the JSON output of the final stage, if applicable.
Token Usage	`token_usage`	`Dict[str, UsageMetrics]`	A summary of token usage across all stages of the pipeline kickoff.
Trace	`trace`	`List[Any]`	A trace of the journey of inputs through the pipeline kickoff.
Crews Outputs	`crews_outputs`	`List[CrewOutput]`	A list of `CrewOutput` objects, representing the outputs from each crew in the pipeline kickoff.

Pipeline Run Result Methods and Properties¶

Method/Property	Description
json	Returns the JSON string representation of the run result if the output format of the final task is JSON.
to_dict	Converts the JSON and Pydantic outputs to a dictionary.
str	Returns the string representation of the run result, prioritizing Pydantic, then JSON, then raw.

Accessing Pipeline Outputs¶

Once a pipeline has been executed, its output can be accessed through the PipelineOutput object returned by the process_runs method. The PipelineOutput class provides access to individual PipelineRunResult objects, each representing a single run through the pipeline.

Example¶

# Define input data for the pipeline
input_data = [{"initial_query": "Latest advancements in AI"}, {"initial_query": "Future of robotics"}]

# Execute the pipeline
pipeline_output = await my_pipeline.process_runs(input_data)

# Access the results
for run_result in pipeline_output.run_results:
    print(f"Run ID: {run_result.id}")
    print(f"Final Raw Output: {run_result.raw}")
    if run_result.json_dict:
        print(f"JSON Output: {json.dumps(run_result.json_dict, indent=2)}")
    if run_result.pydantic:
        print(f"Pydantic Output: {run_result.pydantic}")
    print(f"Token Usage: {run_result.token_usage}")
    print(f"Trace: {run_result.trace}")
    print("Crew Outputs:")
    for crew_output in run_result.crews_outputs:
        print(f"  Crew: {crew_output.raw}")
    print("\n")

This example demonstrates how to access and work with the pipeline output, including individual run results and their associated data.

Using Pipelines¶

Pipelines are particularly useful for complex workflows that involve multiple stages of processing, analysis, or content generation. They allow you to:

Sequence Operations: Execute crews in a specific order, ensuring that the output of one crew is available as input to the next.
Parallel Processing: Run multiple crews concurrently within a stage for increased efficiency.
Manage Complex Workflows: Break down large tasks into smaller, manageable steps executed by specialized crews.

Example: Running a Pipeline¶

# Define input data for the pipeline
input_data = [{"initial_query": "Latest advancements in AI"}]

# Execute the pipeline, initiating a run for each input
results = await my_pipeline.process_runs(input_data)

# Access the results
for result in results:
    print(f"Final Output: {result.raw}")
    print(f"Token Usage: {result.token_usage}")
    print(f"Trace: {result.trace}")  # Shows the path of the input through all stages

Advanced Features¶

Parallel Execution within Stages¶

You can define parallel execution within a stage by providing a list of crews, creating multiple branches:

parallel_analysis_crew = Crew(agents=[financial_analyst], tasks=[financial_analysis_task])
market_analysis_crew = Crew(agents=[market_analyst], tasks=[market_analysis_task])

my_pipeline = Pipeline(
    stages=[
        research_crew,
        [parallel_analysis_crew, market_analysis_crew],  # Parallel execution (branching)
        writing_crew
    ]
)

Routers in Pipelines¶

Routers are a powerful feature in crewAI pipelines that allow for dynamic decision-making and branching within your workflow. They enable you to direct the flow of execution based on specific conditions or criteria, making your pipelines more flexible and adaptive.

What is a Router?¶

A router in crewAI is a special component that can be included as a stage in your pipeline. It evaluates the input data and determines which path the execution should take next. This allows for conditional branching in your pipeline, where different crews or sub-pipelines can be executed based on the router's decision.

Key Components of a Router¶

Routes: A dictionary of named routes, each associated with a condition and a pipeline to execute if the condition is met.
Default Route: A fallback pipeline that is executed if none of the defined route conditions are met.

Creating a Router¶

Here's an example of how to create a router:

from crewai import Router, Route, Pipeline, Crew, Agent, Task

# Define your agents
classifier = Agent(name="Classifier", role="Email Classifier")
urgent_handler = Agent(name="Urgent Handler", role="Urgent Email Processor")
normal_handler = Agent(name="Normal Handler", role="Normal Email Processor")

# Define your tasks
classify_task = Task(description="Classify the email based on its content and metadata.")
urgent_task = Task(description="Process and respond to urgent email quickly.")
normal_task = Task(description="Process and respond to normal email thoroughly.")

# Define your crews
classification_crew = Crew(agents=[classifier], tasks=[classify_task]) # classify email between high and low urgency 1-10
urgent_crew = Crew(agents=[urgent_handler], tasks=[urgent_task])
normal_crew = Crew(agents=[normal_handler], tasks=[normal_task])

# Create pipelines for different urgency levels
urgent_pipeline = Pipeline(stages=[urgent_crew])
normal_pipeline = Pipeline(stages=[normal_crew])

# Create a router
email_router = Router(
    routes={
        "high_urgency": Route(
            condition=lambda x: x.get("urgency_score", 0) > 7,
            pipeline=urgent_pipeline
        ),
        "low_urgency": Route(
            condition=lambda x: x.get("urgency_score", 0) <= 7,
            pipeline=normal_pipeline
        )
    },
    default=Pipeline(stages=[normal_pipeline])  # Default to just normal if no urgency score
)

# Use the router in a main pipeline
main_pipeline = Pipeline(stages=[classification_crew, email_router])

inputs = [{"email": "..."}, {"email": "..."}]  # List of email data

main_pipeline.kickoff(inputs=inputs=inputs)

In this example, the router decides between an urgent pipeline and a normal pipeline based on the urgency score of the email. If the urgency score is greater than 7, it routes to the urgent pipeline; otherwise, it uses the normal pipeline. If the input doesn't include an urgency score, it defaults to just the classification crew.

Benefits of Using Routers¶

Dynamic Workflow: Adapt your pipeline's behavior based on input characteristics or intermediate results.
Efficiency: Route urgent tasks to quicker processes, reserving more thorough pipelines for less time-sensitive inputs.
Flexibility: Easily modify or extend your pipeline's logic without changing the core structure.
Scalability: Handle a wide range of email types and urgency levels with a single pipeline structure.

Error Handling and Validation¶

The Pipeline class includes validation mechanisms to ensure the robustness of the pipeline structure:

Validates that stages contain only Crew instances or lists of Crew instances.
Prevents double nesting of stages to maintain a clear structure.