skatepark-lib 0.8.0

Python framework for AI workflows and pipelines.

Skatepark

Skatepark is a Python framework for creating AI workflow DAGs and pipelines. It augments transformer models with tools for accessing external APIs, such as searches, calculators, spreadsheets, docs, email, and many others. Our initial focus is on supporting large language models (LLMs) but we plan to expand framework's capabilities to cover text-to-anything functionality soon.

With Skatepark, you can accomplish the following:

1. 🚰 Build sequential AI pipelines and sprawling DAG workflows for complex use cases.
2. 🧰️ Augment LLMs with chain of thought capabilities and external tools, such as calculators, web search, spreadsheet editors, and API connectors.
3. 💾 Add memory to AI pipelines for context preservation and summarization.

Please note that Skatepark is in early development. Its APIs and documentation are subject to change. For now, this README file is the most accurate source of documentation and examples.

Getting Started

First, install Skatepark with pip:

pip install skatepark

Second, configure an OpenAI client by getting an API key and adding it to your environment as OPENAI_API_KEY. Skatepark uses OpenAI Completions API to execute LLM prompts and to work with LlamaIndex data structures.

With Skatepark, you can create structures, such as Pipelines and Workflows, that are composed of different types of steps. You can also define structures as JSON objects and load them into Skatepark dynamically. Let's start with defining a simple pipeline.

🚰 AI Pipelines and Workflows

Pipelines are lists of steps that are executed sequentially. Pipelines can have Memory, which makes them ideal for storing LLM conversations.

from skatepark import utils
from skatepark.memory import PipelineMemory
from skatepark.steps import PromptStep
from skatepark.structures import Pipeline

pipeline = Pipeline(
    memory=PipelineMemory()
)

pipeline.add_steps(
    PromptStep("{{ args[0] }}"),
    PromptStep("Say the following like a pirate: {{ input }}")
)

pipeline.run("I am Scotty, who are you?")
pipeline.run("Who am I?")

print(utils.Conversation(pipeline.memory).to_string())

Boom! Our first conversation, à la ChatGPT, is here:

Q: I am Scotty, who are you?
A: Arrr, I be an AI language model designed to assist and answer yer questions, matey!
Q: Who am I?
A: Yarrr, ye just introduced yerself as Scotty, so ye be Scotty, matey!

You can dynamically pass arguments to the prompt by using Jinja templates:

PromptStep("tell me about {{ topic }}", context={"topic": "the hobbit novel"})

In addition to user-defined fields, the context object contains the following:

As Part of Pipeline

• args: arguments passed to the Construct.run() method.
• input: input from the parent.
• structure: the structure that the step belongs to.
• parent: parent step.
• child: child step.

As Part of Workflow

• args: arguments passed to the Construct.run() method.
• inputs: inputs into the current step referencable by parent step IDs.
• structure: the structure that the step belongs to.
• parents: parent steps referencable by IDs.
• children: child steps referencable by IDs.

Skatepark uses OpenAI's gpt-3.5-turbo model by default. If you want to use a different model, set a custom OpenAI prompt driver:

Pipeline(
    prompt_driver=OpenAiPromptDriver(temperature=0.1, model="gpt-4")
)

Now, let's build a simple workflow. Let's say, we want to write a story in a fantasy world with some unique characters. We could setup a workflow that generates a world based on some keywords. Then we pass the world description to any number of child steps that create characters. Finally, the last step pulls in information from all parent steps and writes up a short story.

def character_step(step_id, character_name) -> PromptStep:
    return PromptStep(
        "Based on the following world description create a character named {{ name }}:\n{{ inputs['world'] }}",
        context={
            "name": character_name
        },
        id=step_id
    )

world_step = PromptStep(
    "Create a fictional world based on the following key words {{ keywords|join(', ') }}",
    context={
        "keywords": ["fantasy", "ocean", "tidal lock"]
    },
    id="world"
)

character_step_1 = character_step("scotty", "Scotty")
character_step_2 = character_step("annie", "Annie")

story_step = PromptStep(
    "Based on the following description of the world and characters, write a short story:\n{{ inputs['world'] }}\n{{ inputs['scotty'] }}\n{{ inputs['annie'] }}",
    id="story"
)

workflow = Workflow()

workflow.add_step(world_step)

world_step.add_child(character_step_1)
world_step.add_child(character_step_2)
world_step.add_child(story_step)

character_step_1.add_child(story_step)
character_step_2.add_child(story_step)

workflow.run()

[print(step.output.value) for step in workflow.output_steps()]

And here is the beginning of our story:

Scotty and Annie had been friends since childhood, and their bond had only grown stronger over the years. Scotty had always been fascinated by the ocean and its secrets, and Annie had always been drawn to its magical creatures. [...]

Workflows and pipelines can also be defined in JSON files and loaded dynamically in Python:

{
  "prompt_driver": {
    "temperature": 0.5,
    "type": "OpenAiPromptDriver"
  },
  "steps": [
    {
      "id": "world",
      "type": "PromptStep",
      "parent_ids": [],
      "child_ids": [
        "scotty",
        "annie"
      ],
      "prompt_template": "Create a fictional world based on the following key words {{ keywords|join(', ') }}",
      "context": {
        "keywords": [
          "fantasy",
          "ocean",
          "tidal lock"
        ]
      }
    },
    ...
  ]
}

Here is how you can load and run it:

with open("workflow.json", "r") as file:
    workflow = Workflow.from_json(file.read())

    workflow.run()

🧰️ Tools

The most powerful feature of Skatepark is the ability of workflow and pipeline prompt steps to generate chains of thought and use tools that can interact with the outside world. We use the ReAct technique to implement reasoning and acting in the underlying LLMs without using any fine-tuning. There are two types of tool steps that Skatepark supports:

• ToolStep takes one tool as a parameter and passes it to the LLM that decides if it should use it to respond to the prompt.
• ToolkitStep takes multiple tools as a parameter, so that the underlying LLM can decide which tool to use for every chain of thought step.

Here is how to use tools:

pipeline = Pipeline()

pipeline.add_steps(
    ToolStep(
        "Research and summarize the most important events of February 2023",
        tool=WikiTool()
    ),
    ToolkitStep(
        "Calculate 3^12 and send an email with the answer and the following text to [e-mail]:\n{{ input }}",
        tools=[
            CalculatorTool(),
            EmailSenderTool(
                host="localhost",
                port=1025,
                from_email="[e-mail]",
                use_ssl=False
            )
        ],
        id="calc_email"
    )
)

pipeline.run()

ToolStep instructs an LLM to use a WikiTool that provides a JSON schema and few-shot learning examples that the LLM is automatically "trained" on to interact with Skatepark. The LLM can then decide to use a tool to provide a better prompt response by adding substeps that follow the Thought/Action/Observation ReAct routine. For this prompt, it can obviously use a Wiki tool to obtain new information.

ToolkitStep works the same way, but it provides multiple tools for the LLM to choose from depending on the task. In our example, the LLM uses CalculatorTool to calculate 3^12 and EmailSenderTool to send an email.

Skatepark supports multiple tools and allows you to implement your own.

AwsTool

This tool enables LLMs to run AWS CLI commands. Before using this tool, make sure to install and configure AWS CLI v2.

ToolStep(
    "show me all of my VPCs",
    tool=AwsTool()
)

Warning By default, this tool uses CommandRunner, which executes commands locally in a subprocess. This is not ideal for production environments, where you generally want to execute arbitrary commands in a container. We are working on adding more command runner options soon.

CalculatorTool

This tool enables LLMs to make simple calculations. Here's how to use it:

ToolStep(
    "what's 123^321?",
    tool=CalculatorTool()
)

The LLM will be prompted to reason via the Thought/Action/Observation loop to use the calculator and respond with an answer that the calculator provided.

Warning By default, this tool uses PythonRunner, which executes code locally with sanitized exec. This is not ideal for production environments, where you generally want to execute arbitrary code in a container. We are working on adding more code runner options soon.

SqlClientTool

This tool enables LLMs to execute SQL statements via SQLAlchemy. Depending on your underlying SQL engine, configure your engine_url and give the LLM a hint about what engine you are using via engine_hint, so that it can create engine-specific statements.

ToolStep(
    "list the last 20 items in the orders table",
    tool=SqlClientTool(
        engine_url="sqlite:///skatepark.db",
        engine_hint="sqlite"
    )
)

DataScientistTool

This tool enables LLMs to run more complex calculations in Python. The user can notify the LLM which libraries are available by specifying them in the constructor. By default, only math is available.

ToolStep(
    "what's 123^321?",
    tool=DataScientistTool(
        libs={"numpy": "np", "math": "math"}
    )
)

This will make numpy available as np via import numpy as np and math as math via import math. Before injecting libraries in the constructor, make sure they are installed in your current environment.

Warning By default, this tool uses PythonRunner, which executes code locally with sanitized exec. This is not ideal for production environments, where you generally want to execute arbitrary code in a container. We are working on adding more code runner options soon.

GoogleSearchTool

This tool enables LLMs to search Google. Every search returns links, titles, and short descriptions. Search has two modes: scraping (default) and API-based. To enable API-based search set use_api, api_search_key, and api_search_id params.

ToolStep(
    "Find the latest on LLMs",
    tool=GoogleSearchTool()
)

GoogleSheetsWriterTool and GoogleSheetsReaderTool

These tools enable LLMs to read from and write to Google Sheets worksheets. Before using those tools, make sure to download the service account credentials JSON file and share your spreadsheet with the service account. For more information refer to the gspread auth docs.

To read from a spreadsheet:

ToolStep(
    "read all spreadsheet values from the 2nd and 3rd columns",
    tool=GoogleSheetsReaderTool(
        auth_key_path=os.path.expanduser("~/Desktop/service_account.json"),
        spreadsheet_key="<Google Sheets spreadsheet ID>",
        worksheet_name="<optional worksheet name, defaults to the first worksheet>"
    )
)

To write to a spreadsheet:

ToolStep(
    "Create a spreadsheet with columns for 2022 months in the MM/YYYY format, last column for totals, and rows for profit, revenue, and loss",
    tool=GoogleSheetsWriterTool(
        auth_key_path=os.path.expanduser("~/Desktop/service_account.json"),
        spreadsheet_key="<Google Sheets spreadsheet ID>",
        worksheet_name="<optional worksheet name, defaults to the first worksheet>"
    )
)

EmailSenderTool

This tool enables LLMs to send emails.

ToolStep(
    "send an email with a haiku to [e-mail]",
    EmailSenderTool(
        host="localhost",
        port=1025,
        from_email="[e-mail]",
        use_ssl=False
    )
)

For debugging purposes, you can run a local SMTP server that the LLM will send emails to:

python -m smtpd -c DebuggingServer -n localhost:1025

User the WARPSPEED_EMAIL_SENDER_TOOL_PASSWORD environment variable to set the password.

WebScraperTool

This tool enables LLMs to scrape web pages for full text, summaries, authors, titles, and keywords. It can also execute search queries to answer specific questions about the page.

ToolStep(
    "Can you tell me what's on this page? https://github.com/griptape-ai/skatepark",
    tool=WebScraperTool()
)

WikiTool

This tool enables LLMs to search and query Wikipedia articles:

ToolStep(
    "Research and summarize biggest world news stories in February of 2023",
    tool=WikiTool()
)

Building Your Own Tool

Building your own tools is easy with Skatepark! All you need is a Python class, JSON schema to describe tool actions to the LLM, a set of examples, and a Marshmallow schema for serialization/deserialization. Let's walk through all the required steps and build a simple random number generator tool.

First, create a Python class in a separate directory that generates a random float and optionally truncates it:

import random
from typing import Optional
from skatepark.tools import Tool


class RandomGenTool(Tool):
    def run(self, num_of_decimals: Optional[int]) -> float:
        if num_of_decimals is None:
            return random.random()
        else:
            return round(random.random(), num_of_decimals)

Add a schema.json file describing the tool:

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "random_gen",
  "description": "This tool can be used to generate random numbers",
  "type": "object",
  "properties": {
    "tool": {
      "type": "string",
      "enum": ["random_gen"]
    },
    "input": {
      "type": "int",
      "description": "The number of decimals to be considered while rounding. Default to null."
    }
  },
  "required": ["tool", "input"]
}

Finally, add an examples.j2 Jinja file with a couple of few-shot learning examples:

Input: generate a random number
Thought: I need to use the random_gen tool to answer this question.
Action: {"tool": "random_gen", "input": null}
Observation: 0.8444218515250481
Thought: I have enough information to answer the question
Output: 0.8444218515250481

Input: generate a random number and round it to 2 decimal places
Thought: I need to use the random_gen tool to answer this question.
Action: {"tool": "random_gen", "input": 2}
Observation: 0.14
Thought: I have enough information to answer the question
Output: 0.14

Finally, if you want to use to_json and from_json serialization/deserialization methods, you'll have to add a Marshmallow schema to your tool:

from marshmallow import post_load
from skatepark.schemas import BaseSchema


class RandomGenToolSchema(BaseSchema):
    @post_load
    def make_obj(self, data, **kwargs):
        from .random_gen.random_gen_tool import RandomGenTool

        return RandomGenTool(**data)

The schema class has to be in the following format: <ToolClassName>Schema and be located in <tool_class_name>_schema.py.

Before using the tool, make sure to create an __init__.py file in the tool directory with the following contents:

from .random_gen_tool import RandomGenTool
from .random_gen_tool_schema import RandomGenToolSchema

__all__ = [
    "RandomGenTool",
    "RandomGenToolSchema"
]

Finally, to use the tool:

from skatepark.steps import ToolStep
from random_gen.random_gen_tool import RandomGenTool


ToolStep(
    "generate a random number and round it to 3 decimal places",
    tool=RandomGenTool()
)

If you are deserializing a workflow or a pipeline from JSON, make sure to specify deserialization schema namespace:

{
  "schema_namespace": "random_gen.random_gen_tool_schema",
  "type": "RandomGenTool"
}

Check out other Skatepark tools to learn more about tools' implementation details.

💾 Memory

Skatepark supports different types of memory for pipelines. Due to the non-linear nature of workflows you can't use memory with them yet, but we are currently investigating other possibilities.

By default, pipelines don't initialize memory, so you have to explicitly pass it to them:

Pipeline(
    memory=PipelineMemory()
)

There are two other types of memory: BufferPipelineMemory and SummaryPipelineMemory. BufferPipelineMemory will keep a sliding window of steps that are used to construct a prompt:

Pipeline(
    memory=BufferPipelineMemory(buffer_size=3)
)

This works great for shorter pipelines but fails if the whole workflow context needs to be present. You can use SummaryMemory to address that:

Pipeline(
    memory=SummaryPipelineMemory(
        summarizer=PromptDriverSummarizer(
            driver=OpenAiPromptDriver()
        ),
        offset=2
    )
)

This will progressively summarize the whole pipeline except for the last two steps.

Finally, you can persist memory by using memory drivers. Skatepark comes with one memory driver for automatically storing memory in a file on the disk. Here is how you can initialize memory with a driver:

PipelineMemory(
    driver=DiskMemoryDriver(file_path="memory.json")
)

To load memory:

DiskMemoryDriver(file_path="memory.json").load()

You can easily build drivers for your own data stores by extending MemoryDriver. You only need to implement store and load methods.

Running Tests

Before running tests:

1. git clone https://github.com/griptape-ai/skatepark
2. pip install -r requirements-dev.txt
3. pip install skatepark
4. cd skatepark

To run unit tests: pytest tests/unit/

To run the tests with coverage: pytest --cov --cov-report html tests/unit/

To see the coverage report go to skatepark/htmlcov/index.html

Contributing

Contributions in the form of bug reports, feature ideas, or pull requests are super welcome! Take a look at the current issues and if you'd like to help please submit a pull request with some tests.

License

Skatepark is available under the Apache 2.0 License.