First Workflow
Overview
In this section we're going to create the first workflow in our Movie Comparison application. By the end of this section you will have a workflow that lets you create a cursory comparison between two movies.
For example, if you pass it these two descriptions:
- "A boy discovers an alien in his back yard"
- "A shark attacks a beach"
It will come back with:
Result
Both E.T. the Extra-Terrestrial and Jaws are directed by Steven Spielberg.
They are also both iconic films that have had a significant impact on popular culture.
The hierarchy we will create looks like the following:
graph TB
subgraph " "
direction TB
B(Describe Workflow):::main --> C(Movie 1) --> E(Compare Task):::main
B --> D(Movie 2) --> E
end
classDef main fill:#4274ff1a, stroke:#426effAs you can see, there is a Workflow Structure, and four tasks that will be created. The Start and End tasks (Describe Workflow and Compare Task), and two siblings (Movie 1 and Movie 2). Compare Task will be dependent on them to complete before it can execute.
Tip
Workflows must always have a Start and End task.
To generate this structure, we will first create our Start and End tasks, and then insert the movie tasks. This will guarantee that our tasks exist in the workflow exactly where we want them.
flowchart TB
subgraph Step 2 ["Insert"]
direction TB
C(Describe Workflow):::main --> D(Movie 1) --> F(Compare Task):::main
C --> E(Movie 2) --> F
end
subgraph Step 1 ["Start & End"]
direction TB
A(Describe Workflow):::main --> B(Compare Task):::main
end
classDef main fill:#4274ff1a, stroke:#426effLet's get started.
Importing Required Modules
Before starting, we need to import the necessary modules. Open the app.py file you created in the setup section and import the two Griptape classes you'll need: Workflow and PromptTask:
Note
You might recall that Agent was also imported through griptape.structures. That's because Agent, Workflow, and Pipeline are all Griptape's ways of working with LLMs.
Create Workflow Structure
Initialize the Workflow
Now, let's create the foundation for our Workflow. After the line load_dotenv(), create an instance of the Workflow class:
# ... truncated for brevity
load_dotenv() # Load your environment
# Create the workflow object
workflow = Workflow()
Create a Start task
First, we'll create our "start" task. This will be a simple PromptTask that lets the LLM know what we're going to do.
After the workflow line, add:
# Create tasks
start_task = PromptTask("I will provide you a list of movies to compare.", id="START")
There are two things that are important to point out in this task creation.
- Notice that we're using a
PromptTask. That's because we don't need to use any tools for this particular task, we only want to use the LLM. - We have given the task an
id. This is so we can reference it later in the script. If you don't pass this value, the task will be given a random string as the id.
Important
Every task in a workflow must have a unique id. If two tasks have the same id, the workflow will fail.
Add Task to the Workflow
You have created the task, but it's not yet part of the workflow. In order to do that, we'll need to use the add_task method.
After the PromptTask line, add:
At this point, your workflow flow graph looks like:
graph TB
subgraph " "
direction TB
B(PromptTask: START):::main
end
classDef main fill:#4274ff1a, stroke:#426effHowever if you execute your script, nothing will happen. That's because you need to tell the workflow graph to run.
Run the Workflow
To run a workflow, you simply need to call the method run:
Here's the result. Notice in the logs you can see the Task inputs and outputs:
[11/15/23 13:05:49] INFO PromptTask START
Input: I will provide you a list of movies to compare.
[11/15/23 13:05:52] INFO PromptTask START
Output: Sure, please provide the list of movies you want me to compare.
As you can see, the LLM is ready to take a list of movies.
Add the end task
The next task we'll add will be the last task. This is a good place to create a task that summarizes what has been done, and then add it to the end of the workflow using the same add_task method.
# Create tasks
start_task = PromptTask("I will provide you a list of movies to compare.", id="START")
end_task = PromptTask("How are the movies the same?", id="END")
# Add tasks to workflow
workflow.add_task(start_task)
workflow.add_task(end_task)
graph TB
subgraph " "
direction TB
B(PromptTask: START):::main --> C(PromptTask: END):::main
end
classDef main fill:#4274ff1a, stroke:#426effTip
This process has changed from versions of Griptape prior to v0.20.
In previous versions, the add_task method would add tasks as siblings of the parent task. With versions greater than 0.20, they add them one after another. To add tasks as siblings you will be inserting tasks.
graph TB
subgraph " "
direction TB
B(PromptTask: START):::main --> C(PromptTask: END):::main
end
classDef main fill:#4274ff1a, stroke:#426effgraph TB
subgraph " "
direction TB
B(PromptTask: START):::main
C(PromptTask: END):::main
end
classDef main fill:#4274ff1a, stroke:#426effAdd Tasks into Workflow
Create the first Movie Prompt task
Now it's time to create the first task asking the LLM to identify a movie. This will be a PromptTask.
In your code, after the end_task, add:
# Create tasks
start_task = PromptTask("I will provide you a list of movies to compare.", id="START")
end_task = PromptTask("How are the movies the same?", id="END")
# Create movie tasks
movie_1_task = PromptTask("What movie is this: boy finds alien in backyard.", id="movie_1")
Insert the Task into the Workflow
You have created the task, but it's not yet part of the workflow. In order to do that, we'll need to use the insert_tasks method.
The insert_tasks method takes a couple of arguments. It looks like this:
workflow.insert_tasks(task_a, [task_c, task_d], task_b)
workflow.insert_tasks(task_c, [task_e, task_f], task_b)
Notice the second argument is a list of tasks. This method will take two tasks that already have a parent/child relationship, and insert whatever tasks are listed in that second argument between them.
The graph above would look something like:
graph TB
subgraph " "
direction TB
A(task_a)
B(task_b)
C(task_c)
D(task_d)
E(task_e)
F(task_f)
A --> B --> D
A --> C --> E --> D
C --> F --> D
end
We'll see this in more detail shortly, but for now let's just add the single task.
In your script, after the add_task lines, add:
At this point, your workflow flow graph looks like:
graph TB
subgraph " "
direction TB
A(PromptTask: START):::main
B(PromptTask: movie_1_task)
C(PromptTask: END):::main
A --> B --> C
end
classDef main fill:#4274ff1a, stroke:#426effRun the Workflow
Run the script to see the output.
Here's the result. Notice the tasks being executed, and the summary where it tries to compare the results:
[11/15/23 14:10:59] INFO PromptTask START
Input: I will provide you a list of movies to compare.
[11/15/23 14:11:00] INFO PromptTask START
Output: Sure, please provide the list of movies you want me to compare.
INFO PromptTask movie_1
Input: What movie is this: boy finds alien in backyard.
[11/15/23 14:11:02] INFO PromptTask movie_1
Output: The movie you are referring to is likely "E.T. the Extra-Terrestrial" directed by Steven
Spielberg.
INFO PromptTask END
Input: How are the movies the same?
[11/15/23 14:11:06] INFO PromptTask END
Output: As an AI, I need specific movies to compare in order to provide similarities. Please
provide the names of the movies you want to compare.
The LLM is unable to compare the results for two reasons - first, we only passed a single movie. Second, we didn't pass the information about the movie back to the LLM so it doesn't know what we were talking about. We'll take care of these one at a time.
First, we'll add a second movie.
Second Movie Task
Just below the first movie PromptTask, add a second one with another description.
# ...
# Create tasks
movie_1_task = PromptTask("What movie is this: boy finds alien in backyard.", id="movie_1")
movie_2_task = PromptTask("What movie is this?: a shark attacks a beach.", id="movie_2")
# ...
Important
Don't forget to add the id to the second task and make sure it's unique from the first task.
Add Second Task to Workflow
Just like before, we need to insert movie_2_task to workflow with the insert_tasks method. Modify the previous insert_tasks line to include both movie tasks.
# ...
# Add tasks to workflow
workflow.insert_tasks(start_task, [movie_1_task, movie_2_task], end_task)
# ...
Test
Go ahead and run the code and notice that the two movie tasks executed in parallel:
# ... truncated for brevity
[11/15/23 14:14:32] INFO PromptTask START
Input: I will provide you a list of movies to compare.
[11/15/23 14:14:34] INFO PromptTask START
Output: Sure, I am ready to help you compare the movies. Please provide the list.
INFO PromptTask movie_2
Input: What movie is this?: a shark attacks a beach.
INFO PromptTask movie_1
Input: What movie is this: boy finds alien in backyard.
[11/15/23 14:14:38] INFO PromptTask movie_2
Output: This could refer to several movies as shark attacks are a common theme in many films.
However, the most iconic one is "Jaws" directed by Steven Spielberg.
INFO PromptTask movie_1
Output: This could refer to several movies, but the most famous one is probably "E.T. the
Extra-Terrestrial" directed by Steven Spielberg.
INFO PromptTask END
Input: How are the movies the same?
[11/15/23 14:14:41] INFO PromptTask END
Output: As an AI, I need more specific details to provide a comparison. Please provide the names of
the movies you want to compare.
Hmm. It doesn't look like the compare task knows what we're talking about. The workflow evaluated both we sent it, we can see that in the logs above, but the compare task has no knowledge of them.
That's because we need to pass the results of the previous task to the current task. This is a very important feature, as it allows us to be very specific about what data is sent to the LLM.
Pass the Data
In order to send the data to the PromptTask, we need to somehow feed the result of the previous task's execution to the prompt.
Jinja2
Griptape uses the Jinja2 template engine, which allows you to insert data into the prompt. There's a lot of power available with Jinja templates, but in this course we'll keep our focus rather small.
Jinja templates access variables using the {{ }} syntax. Tasks have a property parent_outputs that tell us what objects are coming into the node.
Update the Prompt
In order to update the prompt, we want to tell it what tasks to be looking at. If you check your code, you can see that we used the id property earlier when we were creating the tasks:
movie_1_task = PromptTask("What movie is this?: A boy discovers an alien in his back yard", id="movie_1")
movie_2_task = PromptTask("What movie is this?: a shark attacks a beach.", id="movie_2")
So movie_1 and movie_2 are the two ids we can use in our Jinja template. They can be specified like this: {{ parent_outputs['movie_1'] }} and {{ parent_outputs['movie_2'] }}
Update the END task to specify the particular ids of the parent_outputs. Note - I'm using """ in order to allow us to use multiple lines for the PromptTask string.
end_task = PromptTask("""
How are these movies the same:
{{ parent_outputs['movie_1'] }}
{{ parent_outputs['movie_2'] }}
""",
id="END")
Test
When you run the script now, you should see a much better Input:
[11/15/23 14:19:24] INFO PromptTask END
Input:
How are these movies the same:
This could be referring to "E.T. the Extra-Terrestrial" directed by Steven Spielberg.
This could be referring to several movies as shark attacks are a common theme in films.
However, the most iconic movie featuring a shark attack on a beach is "Jaws" directed by Steven
Spielberg.
[11/15/23 14:19:26] INFO PromptTask END
Output: These movies are the same in that they are both directed by Steven Spielberg.
The code works but it's if you look closely at the END task input, you'll see that it's using the full string of the response from the prompts:
How are these movies the same:
That sounds like the movie "E.T. the Extra-Terrestrial".
That could be several movies, but the most famous one is probably "Jaws".
What we really want is just the name of the movie, not any commentary. We can fix that by adjusting the prompt in the initial query.
Fix the Prompt
Change the movie PromptTasks so we ask for just the name:
movie_1_task = PromptTask(
"What movie is this? Return only the movie name: A boy discovers an alien in his back yard",
id="movie_1")
movie_2_task = PromptTask(
"What movie is this? Return only the movie name: a shark attacks a beach.",
id="movie_2")
After running the response input task should be much cleaner:
Code Review
We covered quite a lot of ground creating your first workflow. Double-check your script and make sure you've got it working as expected:
Next Step
In the next section we are going to make our script a bit more flexible by making it possible to compare as many movie descriptions as we want. Check out Adding Flexibility when you're ready to continue.