Overview
The Griptape framework provides developers with the ability to create AI systems that operate across two dimensions: predictability and creativity.
For predictability, Griptape enforces structures like sequential pipelines, DAG-based workflows, and long-term memory. To facilitate creativity, Griptape safely prompts LLMs with tools and short-term memory connecting them to external APIs and data stores. The framework allows developers to transition between those two dimensions effortlessly based on their use case.
Griptape not only helps developers harness the potential of LLMs but also enforces trust boundaries, schema validation, and tool activity-level permissions. By doing so, Griptape maximizes LLMs’ reasoning while adhering to strict policies regarding their capabilities.
Griptape’s design philosophy is based on the following tenets:
- Modularity and composability: All framework primitives are useful and usable on their own in addition to being easy to plug into each other.
- Technology-agnostic: Griptape is designed to work with any capable LLM, data store, and backend through the abstraction of drivers.
- Keep data off prompt by default: When working with data through loaders and tools, Griptape aims to keep it off prompt by default, making it easy to work with big data securely and with low latency.
- Minimal prompt engineering: It’s much easier to reason about code written in Python, not natural languages. Griptape aims to default to Python in most cases unless absolutely necessary.
Quick Start
OpenAI API Key
First, configure an OpenAI client by getting an API key and adding it to your environment as OPENAI_API_KEY.
By default, Griptape uses OpenAI Completions API to execute LLM prompts, but other LLMs can be configured with the use of Prompt Drivers.
Using pip
Install griptape:
Using Poetry
To get started with Griptape using Poetry first create a new poetry project from the terminal:
Change your working directory to the new griptape-quickstart directory created by Poetry and add the the griptape dependency.
Extras
The [all] extra ensures that you have access to the entire range of functionalities that Griptape offers.
This comprehensive installation is recommended for newcomers to get the complete Griptape experience.
However, if you wish to optimize the installation size or only require specific functionalities, you have two main options:
- Core Dependencies: These are the foundational dependencies that enable Griptape to function with most of its default settings.
- Extras: These are additional, vendor-specific drivers integrated within the Griptape framework. If a particular Driver mandates an extra, it will be explicitly highlighted in the documentation.
To install just the core dependencies:
To install specific extras (e.g., drivers for AnthropicPromptDriver and PineconeVectorStoreDriver):
For a comprehensive list of extras, please refer to the [tool.poetry.extras] section of Griptape's pyproject.toml.
Build a Simple Agent
With Griptape, you can create structures, such as Agents, Pipelines, and Workflows, that are composed of different types of tasks. First, let's build a simple Agent that we can interact with through a chat based interface.
from griptape.structures import Agent
from griptape.utils import Chat
agent = Agent()
Chat(agent).start()
Q: prompt where you can interact with your model.
Q: Write me a haiku about griptape
processing...
[09/08/23 09:52:45] INFO PromptTask d4302227570e4a978ed79e7e0444337b
Input: Write me a haiku about griptape
[09/08/23 09:52:48] INFO PromptTask d4302227570e4a978ed79e7e0444337b
Output: Griptape rough and true,
Skateboard's trusty, silent guide,
In each ride, we're glued.
A: Griptape rough and true,
Skateboard's trusty, silent guide,
In each ride, we're glued.
Q:
.run() method:
from griptape.structures import Agent
from griptape.utils import Chat
agent = Agent()
agent.run("write me a haiku about griptape")
Build a Simple Agent with Tools
from griptape.structures import Agent
from griptape.tools import Calculator
calculator = Calculator()
agent = Agent(
tools=[calculator]
)
agent.run(
"what is 7^12"
)
[09/08/23 09:53:42] INFO ToolkitTask c87320c5ab1b4988acf25c107b46dffa
Input: what is 7^12
[09/08/23 09:53:49] INFO Subtask f3f41104a2234a69832c7eacb64e7324
Thought: The user is asking for the value of 7 raised to the power of 12. I can use the Calculator tool to
perform this calculation.
Action: {"name": "Calculator", "path": "calculate", "input": {"values": {"expression":
"7**12"}}}
INFO Subtask f3f41104a2234a69832c7eacb64e7324
Response: 13841287201
[09/08/23 09:53:51] INFO ToolkitTask c87320c5ab1b4988acf25c107b46dffa
Output: The value of 7 raised to the power of 12 is 13841287201.
Build a Simple Pipeline
Agents are great for getting started, but they are intentionally limited to a single task. Pipelines, however, allow us to define any number of tasks to run in sequence. Let's define a simple two-task Pipeline that uses tools and memory:
from griptape.memory.structure import ConversationMemory
from griptape.structures import Pipeline
from griptape.tasks import ToolkitTask, PromptTask
from griptape.tools import WebScraper, FileManager, TaskMemoryClient
# Pipelines represent sequences of tasks.
pipeline = Pipeline(
conversation_memory=ConversationMemory()
)
pipeline.add_tasks(
# Load up the first argument from `pipeline.run`.
ToolkitTask(
"{{ args[0] }}",
# Add tools for web scraping, and file management
tools=[WebScraper(), FileManager(), TaskMemoryClient(off_prompt=False)]
),
# Augment `input` from the previous task.
PromptTask(
"Say the following in spanish: {{ parent_output }}"
)
)
pipeline.run(
"Load https://www.griptape.ai, summarize it, and store it in griptape.txt"
)
[09/08/23 10:02:34] INFO ToolkitTask 3c1d2f4a49384873820a9a8cd8acc983
Input: Load https://www.griptape.ai, summarize it, and store it in griptape.txt
[09/08/23 10:02:44] INFO Subtask 42fd56ba100e45688401c5ce32b79a33
Thought: To complete this task, I need to first load the webpage using the WebScraper tool's get_content
activity. Then, I will summarize the content using the TaskMemory tool's summarize activity. Finally, I will
store the summarized content in a file named griptape.txt using the FileManager tool's save_file_to_disk
activity.
Action: {"name": "WebScraper", "path": "get_content", "input": {"values": {"url":
"https://www.griptape.ai"}}}
[09/08/23 10:02:45] INFO Subtask 42fd56ba100e45688401c5ce32b79a33
Response: Output of "WebScraper.get_content" was stored in memory with memory_name "TaskMemory" and
artifact_namespace "39ca67bbe26b4e1584193b87ed82170d"
[09/08/23 10:02:53] INFO Subtask 8023e3d257274df29065b22e736faca8
Thought: Now that the webpage content is stored in memory, I can use the TaskMemory tool's summarize activity
to summarize the content.
Action: {"name": "TaskMemoryClient", "path": "summarize", "input": {"values": {"memory_name": "TaskMemory", "artifact_namespace": "39ca67bbe26b4e1584193b87ed82170d"}}}
[09/08/23 10:02:57] INFO Subtask 8023e3d257274df29065b22e736faca8
Response: Griptape is an open source framework that allows developers to build and deploy AI applications
using large language models (LLMs). It provides the ability to create conversational and event-driven apps that
can securely access and manipulate data. The framework enforces structures for predictability and creativity,
allowing developers to easily transition between the two. Griptape Cloud is a managed platform for deploying and
managing AI apps.
[09/08/23 10:03:06] INFO Subtask 7baae700239943c18b5b6b21873f0e13
Thought: Now that I have the summarized content, I can store it in a file named griptape.txt using the
FileManager tool's save_file_to_disk activity.
Action: {"name": "FileManager", "path": "save_file_to_disk", "input": {"values":
{"memory_name": "TaskMemory", "artifact_namespace": "39ca67bbe26b4e1584193b87ed82170d", "path":
"griptape.txt"}}}
INFO Subtask 7baae700239943c18b5b6b21873f0e13
Response: saved successfully
[09/08/23 10:03:14] INFO ToolkitTask 3c1d2f4a49384873820a9a8cd8acc983
Output: The summarized content of the webpage https://www.griptape.ai has been successfully stored in the file
named griptape.txt.
INFO PromptTask 8635925ff23b46f28a740105bd11ca8f
Input: Say the following in spanish: The summarized content of the webpage https://www.griptape.ai has been
successfully stored in the file named griptape.txt.
[09/08/23 10:03:18] INFO PromptTask 8635925ff23b46f28a740105bd11ca8f
Output: El contenido resumido de la página web https://www.griptape.ai se ha almacenado con éxito en el archivo
llamado griptape.txt.
-
In some cases a model might be capable of basic arithmetic. For example, gpt-3.5 returns the correct numeric answer but in an odd format. ↩