Introducing NPi
Wells Wang<w@npi.ai> • 2024/04/30
Background
Since ChatGPT's release, there has been a surge in AI applications designed for natural conversations. However, their practical usefulness is often limited by a lack of automatic action-taking capabilities
The evolving concept of Agent addresses this gap. Agent AI is a class of interactive systems that can perceive visual stimuli, language inputs, and other environmentally grounded data, and can produce meaningful embodied actions1.
Agents are not only going to change how everyone interacts with computers. They’re also going to upend the software industry, bringing about the biggest revolution in computing since we went from typing commands to tapping on icons.
The future of Agents, Bill Gates2.
A major advantage humans hold over other animals is using tools. This also one of the AI Agents' key abilities, noted by Andrew Ng3.
However, building an AI agent with a robust ability to use tools is challenging due to the diversity of tools and the operational overhead involved.
- Low-level primitives, such as fine-grained HTTP APIs or SDKs, lead organizations to repeatedly writing similar code to integrate LLM with different applications.
- Maintaining non-business critical features, like State Management, Availability, and Authorization flows, incurs significant overhead.
- Ensuring the security of AI Agents — making their actions controllable, predictable, and explainable — poses substantial challenges.
Although AI agent development is progressing rapidly, we identified a significant gap in the current LLM stack ecosystem:
the absence of products specifically tailored to address the challenges of Tool Use
faced by AI agent developers.
To fill this void, we created NPi, a platform designed to provide an easy-to-use and reliable functionalities that
enhances the robust tool-use capabilities of AI agents.
What is NPi?
On April 25, we launched NPi (v0.0.1
) on GitHub (opens in a new tab), a free, open-source platform.
NPi provides Tool Use APIs that empower AI agents to operate and interact with various software tools and applications.
You can explore NPi use cases at the NPi Examples (opens in a new tab).
The primary goal of NPi is to offer a unified interface that allows Large Language Models to seamlessly integrate with the existing software and applications ecosystem through function calls. NPi serves as a gateway for these models to access the virtual world.
The core principle of NPi is to focus on in-app planning
. This requires users to break down tasks into single-app
sub-tasks, meaning each task is confined to one application. NPi then interpreters these sub-tasks into a series of
function calls, executing them in a rule-based manner to ensure precise control.
This method, known as divide-and-conquer, is a common strategy for solving complex problems and is central to NPi's designing.
To date, we have implemented the core functionalities of NPi, including:
Out-of-box multimodal Tool use APIs
We provide ready-to-use APIs that allow large language models to interact with applications, demonstrated in the following examples:
from npiai.app.google import Calendar
from npiai.app.github import GitHub
from npiai.app.twitter import Twitter
from openai import OpenAI
# Use case 1: directly interact with the application
github = GitHub()
github.chat("...")
# Use case 2: integrate application with OpenAI
calendar = Calendar()
calendar.chat("...")
client = OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo-0125",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "What's on my calendar today?"},
],
tools=[calendar.as_openai_tool()],
tool_choice="auto", # auto is default, but we'll be explicit
)
# Use cases 3: for non-API friendly cases, a visual-based approach was applied, powered by Playwright and GPT-4V
twitter = Twitter(visual=True)
twitter.chat("what's the @wellswfwang latest post?")
Under the hood, NPi is pre-integrated with specific applications' SDKs or APIs, interpreting the given task into a sequence of function calls.
We continuously monitor changes in these SDKs and APIs to stay aligned with them. This ensures NPi remains up-to-date, relieving you of the burden of tracking these changes yourself.
Multi-agent collaboration
A clean and easy-to-use interface for building multi-agents applications.
from npiai.core import Agent
agent1 = Agent(prompt="...")
agent1.use(Gmail(), Calendar())
agent2 = Agent(prompt="...")
agent2.use(GitHub())
agent3 = Agent.collaborate(agent1, agent2)
agent3.run(task="...")
agent3
acts as a coordinator, orchestrating the operations of agent1
and agent2
. Furthermore, you also can integrate agent with OpenAI.
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model="gpt-3.5-turbo-0125",
messages=[...],
tools=[agent3.as_openai_tool()],
tool_choice="auto", # auto is default, but we'll be explicit
)
Human-in-the-loop(HITL)
A simple and effective way to ensure human involvement in handling sensitive operations appropriately.
This "human in the loop" approach is an essential step in ensuring that language models behave responsibly, generate accurate responses, and align with ethical and safety standards
Large Language Model: Data, Human in the Loop for Fine-Tuning4
For example, in Gmail
app, we pre-set the sending email action
as sensitive, so that each calling of this action needs human
to approve sending.
# These HITL APIs will be released in v0.0.2
from npiai.core.hitl import HITLRequest, HITLResponse, RequestApproved, RequestDenied, Console
from npiai_proto import api_pb2
def human_assist(req: api_pb2.HITLRequest) -> HITLResponse:
console = Console(req) # you can integrate your own workflow
if req.type == api_pb2.ActionType.SAFEGUARD:
result = console.wait()
if result.is_approved():
return RequestApproved
if req.action == api_pb2.ActionType.HUMAN_FEEDBACK:
result = console.wait()
return result.human_message()
return RequestDenied
gmail = Gmail()
gmail.hitl_handler(human_assist)
Additionally, you can easily change this behavior by providing customized configuration.
Minimize operational overhead
Imagine an AI Agent tasked with coordinating meeting times among stakeholders. Initially, the Agent suggests multiple time slots and sends confirmation requests via email to the stakeholders.
While waiting for human responses, delays are not uncommon. Various issues such as unexpected software shutdowns or network errors can arise. Handling these edge cases is a frequent challenge in programming.
NPi is designed to manage these errors, ensuring smooth recovery from such disruptions and relieving developers of the burden of addressing these complex cases directly.
Additionally, APIs fine-tuning, evaluation, and cost-effectiveness are also in our roadmap.
How does NPi work?
NPi has an architecture that consists of two primary components: the Server and the Toolkits. Typically, the developers use the CLI or Web Console to customize NPi, then use NPi SDKs to add tool-use ability into their AI Agents.
The Server has two main responsibilities:
- Management Functionalities: This includes App API management, authorization, and advanced options such as fine-tuning and evaluating function calls.
- Function Calling Runtime: The server interprets single-app task into a sequence of function calls based on in-app planning and executes them. It also handles the persistence of function call states if necessary and manages communications with the client side. This is particularly crucial for cross-agents communication and incorporating human input.
The Toolkits are a collection of tools that designed to enhance the developer experience:
- Multi-language SDKs: We provide SDKs in various programming languages, making it easy for developers to integrate NPi with their agents and enhance their tool-using capabilities.
- CLI: Our command-line tool offers a straightforward interface for managing apps, authorizing, and exploring NPi’s core functionalities.
- Web Console: A web-based interface that enables more intricate interactions with NPi, including calling-memory management, fine-tuning, evaluation, and observability. (Note: This feature has not been released yet.)
What's next?
We are excited to announce the release of NPi, a platform still in its early stages, We are actively working on implementing features that mentioned previous sections.
The mission of NPi is to act as the limbs for large language models, enabling AI Agents the ability of accessing virtual world. Although AI Agents, particularly in utilizing the Tool Use pattern, have not yet been widely adopted, we are optimistic that NPi will accelerate this adoption and bring us closer to Artificial General Intelligence (AGI).
To realize this vision, we are eager to involve more AI Agent developers and build a vibrant community to collaboratively shape the future of NPi.
Explore our development plans on the NPi Roadmap (opens in a new tab). Connect with us on GitHub (opens in a new tab), X.com (opens in a new tab), or join our Discord Community (opens in a new tab). You can also reach out directly to our CEO (opens in a new tab) via X.com.
Your support, insights, and feedback are highly valued and greatly appreciated!
We look forward to meeting you in the NPi community and exploring this exciting future together.