Skip to content

API for building multi-step agent workflows. (An orchestrator for AI agents and workflows.)

License

Notifications You must be signed in to change notification settings

suvekshyabhandari356/julep

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

English | 中文翻译 | 日本語翻訳

julep


Explore Docs · Discord · 𝕏 · LinkedIn

NPM Version   PyPI - Version   Docker Image Version   GitHub License


Note

👨‍💻 Here for the devfest.ai event? Join our Discord and check out the details below.

🌟 Contributors and DevFest.AI Participants (Click to expand)

🌟 Call for Contributors!

We're excited to welcome new contributors to the Julep project! We've created several "good first issues" to help you get started. Here's how you can contribute:

  1. Check out our CONTRIBUTING.md file for guidelines on how to contribute.
  2. Browse our good first issues to find a task that interests you.
  3. If you have any questions or need help, don't hesitate to reach out on our Discord channel.

Your contributions, big or small, are valuable to us. Let's build something amazing together! 🚀

🎉 DevFest.AI October 2024

Exciting news! We're participating in DevFest.AI throughout October 2024! 🗓️

  • Contribute to Julep during this event and get a chance to win awesome Julep merch and swag! 🎁
  • Join developers from around the world in contributing to AI repositories and participating in amazing events.
  • A big thank you to DevFest.AI for organizing this fantastic initiative!

[!TIP] Ready to join the fun? Tweet that you are participating and let's get coding! 🖥️

[!NOTE] Get your API key here.

While we are in beta, you can also reach out on Discord to get rate limits lifted on your API key.

Julep DevFest.AI

📖 Table of Contents

Introduction

Julep is a platform for creating AI agents that remember past interactions and can perform complex tasks. It offers long-term memory and manages multi-step processes.

Julep enables the creation of multi-step tasks incorporating decision-making, loops, parallel processing, and integration with numerous external tools and APIs.

While many AI applications are limited to simple, linear chains of prompts and API calls with minimal branching, Julep is built to handle more complex scenarios.

It supports:

  • Intricate, multi-step processes
  • Dynamic decision-making
  • Parallel execution

Tip

Imagine you want to build an AI agent that can do more than just answer simple questions—it needs to handle complex tasks, remember past interactions, and maybe even use other tools or APIs. That's where Julep comes in.

Quick Example

Imagine a Research AI agent that can do the following:

  1. Take a topic,
  2. Come up with 100 search queries for that topic,
  3. Perform those web searches in parallel,
  4. Summarize the results,
  5. Send the summary to Discord

In Julep, this would be a single task under 80 lines of code and run fully managed all on its own. All of the steps are executed on Julep's own servers and you don't need to lift a finger. Here's a working example:

name: Research Agent

# Optional: Define the input schema for the task
input_schema:
  type: object
  properties:
    topic:
      type: string
      description: The main topic to research

# Define the tools that the agent can use
tools:
- name: web_search
  type: integration
  integration:
    provider: brave
    setup:
      api_key: "YOUR_BRAVE_API_KEY"

- name: discord_webhook
  type: api_call
  api_call:
    url: "YOUR_DISCORD_WEBHOOK_URL"
    method: POST
    headers:
      Content-Type: application/json

# Special variables:
# - inputs: for accessing the input to the task
# - outputs: for accessing the output of previous steps
# - _: for accessing the output of the previous step

# Define the main workflow
main:
- prompt:
    - role: system
      content: >-
        You are a research assistant.
        Generate 100 diverse search queries related to the topic:
        {{inputs[0].topic}}

        Write one query per line.
  unwrap: true

# Evaluate the search queries using a simple python expression
- evaluate:
    search_queries: "_.split('\n')"

# Run the web search in parallel for each query
- over: "_.search_queries"
  map:
    tool: web_search
    arguments:
      query: "_"
  parallelism: 100

# Collect the results from the web search
- evaluate:
    results: "'\n'.join([item.result for item in _])"

# Summarize the results
- prompt:
    - role: system
      content: >
        You are a research summarizer. Create a comprehensive summary of the following research results on the topic {{inputs[0].topic}}.
        The summary should be well-structured, informative, and highlight key findings and insights:
        {{_.results}}
  unwrap: true

# Send the summary to Discord
- tool: discord_webhook
  arguments:
    content: >
      **Research Summary for {{inputs[0].topic}}**

      {{_}}

Tip

Julep is really useful when you want to build AI agents that can maintain context and state over long-term interactions. It's great for designing complex, multi-step workflows and integrating various tools and APIs directly into your agent's processes.

In this example, Julep will automatically manage parallel executions, retry failed steps, resend API requests, and keep the tasks running reliably until completion.

Key Features

  1. 🧠 Persistent AI Agents: Remember context and information over long-term interactions.
  2. 💾 Stateful Sessions: Keep track of past interactions for personalized responses.
  3. 🔄 Multi-Step Tasks: Build complex, multi-step processes with loops and decision-making.
  4. Task Management: Handle long-running tasks that can run indefinitely.
  5. 🛠️ Built-in Tools: Use built-in tools and external APIs in your tasks.
  6. 🔧 Self-Healing: Julep will automatically retry failed steps, resend messages, and generally keep your tasks running smoothly.
  7. 📚 RAG: Use Julep's document store to build a system for retrieving and using your own data.

Julep is ideal for applications that require AI use cases beyond simple prompt-response models.

Why Julep vs. LangChain?

Different Use Cases

Think of LangChain and Julep as tools with different focuses within the AI development stack.

LangChain is great for creating sequences of prompts and managing interactions with AI models. It has a large ecosystem with lots of pre-built integrations, which makes it convenient if you want to get something up and running quickly. LangChain fits well with simple use cases that involve a linear chain of prompts and API calls.

Julep, on the other hand, is more about building persistent AI agents that can remember things over long-term interactions. It shines when you need complex tasks that involve multiple steps, decision-making, and integration with various tools or APIs directly within the agent's process. It's designed from the ground up to manage persistent sessions and complex tasks.

Use Julep if you imagine building a complex AI assistant that needs to:

  • Keep track of user interactions over days or weeks.
  • Perform scheduled tasks, like sending daily summaries or monitoring data sources.
  • Make decisions based on prior interactions or stored data.
  • Interact with multiple external services as part of its task.

Then Julep provides the infrastructure to support all that without you having to build it from scratch.

Different Form Factor

Julep is a platform that includes a language for describing tasks, a server for running those tasks, and an SDK for interacting with the platform. To build something with Julep, you write a description of the task in YAML, and then run the task in the cloud.

Julep is built for heavy-lifting, multi-step, and long-running tasks and there's no limit to how complex the task can be.

LangChain is a library that includes a few tools and a framework for building linear chains of prompts and tools. To build something with LangChain, you typically write Python code that configures and runs the model chains you want to use.

LangChain might be sufficient and quicker to implement for simple use cases that involve a linear chain of prompts and API calls.

In Summary

Use LangChain when you need to manage AI model interactions and prompt sequences in a stateless or short-term context.

Choose Julep when you need a robust framework for stateful agents with advanced task capabilities, persistent sessions, and complex task management.

Installation

To get started with Julep, install it using npm or pip:

npm install @julep/sdk

or

pip install julep

Note

Get your API key here.

While we are in beta, you can also reach out on Discord to get rate limits lifted on your API key.

Tip

💻 Are you a show me the code!™ kind of person? We have created a ton of cookbooks for you to get started with. Check out the cookbooks to browse through examples.

💡 There's also lots of ideas that you can build on top of Julep. Check out the list of ideas to get some inspiration.

Python Quick Start 🐍

Step 1: Create an Agent

import yaml
from julep import Julep # or AsyncJulep

client = Julep(api_key="your_julep_api_key")

agent = client.agents.create(
    name="Storytelling Agent",
    model="gpt-4o",
    about="You are a creative storytelling agent that can craft engaging stories and generate comic panels based on ideas.",
)

# 🛠️ Add an image generation tool (DALL·E) to the agent
client.agents.tools.create(
    agent_id=agent.id,
    name="image_generator",
    description="Use this tool to generate images based on descriptions.",
    integration={
        "provider": "dalle",
        "method": "generate_image",
        "setup": {
            "api_key": "your_openai_api_key",
        },
    },
)

Step 2: Create a Task that generates a story and comic strip

Let's define a multi-step task to create a story and generate a paneled comic strip based on an input idea:

# 📋 Task
# Create a task that takes an idea and creates a story and a 4-panel comic strip
task_yaml = """
name: Story and Comic Creator
description: Create a story based on an idea and generate a 4-panel comic strip illustrating the story.

main:
  # Step 1: Generate a story and outline into 4 panels
  - prompt:
      - role: system
        content: You are {{agent.name}}. {{agent.about}}
      - role: user
        content: >
          Based on the idea '{{_.idea}}', write a short story suitable for a 4-panel comic strip.
          Provide the story and a numbered list of 4 brief descriptions for each panel illustrating key moments in the story.
    unwrap: true

  # Step 2: Extract the panel descriptions and story
  - evaluate:
      story: _.split('1. ')[0].strip()
      panels: re.findall(r'\\d+\\.\\s*(.*?)(?=\\d+\\.\\s*|$)', _)

  # Step 3: Generate images for each panel using the image generator tool
  - foreach:
      in: _.panels
      do:
        tool: image_generator
        arguments:
          description: _

  # Step 4: Generate a catchy title for the story
  - prompt:
      - role: system
        content: You are {{agent.name}}. {{agent.about}}
      - role: user
        content: >
          Based on the story below, generate a catchy title.

          Story: {{outputs[1].story}}
    unwrap: true

  # Step 5: Return the story, the generated images, and the title
  - return:
      title: outputs[3]
      story: outputs[1].story
      comic_panels: "[output.image.url for output in outputs[2]]"
"""

task = client.tasks.create(
    agent_id=agent.id,
    **yaml.safe_load(task_yaml)
)

Step 3: Execute the Task

# 🚀 Execute the task with an input idea
execution = client.executions.create(
    task_id=task.id,
    input={"idea": "A cat who learns to fly"}
)

# 🎉 Watch as the story and comic panels are generated
for transition in client.executions.transitions.stream(execution_id=execution.id):
    print(transition)

# 📦 Once the execution is finished, retrieve the results
result = client.executions.get(execution_id=execution.id)

Step 4: Chat with the Agent

Start an interactive chat session with the agent:

session = client.sessions.create(agent_id=agent.id)

# 💬 Send messages to the agent
while (message := input("Enter a message: ")) != "quit":
    response = client.sessions.chat(
        session_id=session.id,
        message=message,
    )

    print(response)

Tip

You can find the full python example here.

Node.js Quick Start 🟩

Step 1: Create an Agent

import { Julep } from '@julep/sdk';
import yaml from 'js-yaml';

const client = new Julep({ apiKey: 'your_julep_api_key' });

async function createAgent() {
  const agent = await client.agents.create({
    name: "Storytelling Agent",
    model: "gpt-4",
    about: "You are a creative storytelling agent that can craft engaging stories and generate comic panels based on ideas.",
  });

  // 🛠️ Add an image generation tool (DALL·E) to the agent
  await client.agents.tools.create(agent.id, {
    name: "image_generator",
    description: "Use this tool to generate images based on descriptions.",
    integration: {
      provider: "dalle",
      method: "generate_image",
      setup: {
        api_key: "your_openai_api_key",
      },
    },
  });

  return agent;
}

Step 2: Create a Task that generates a story and comic strip

const taskYaml = `
name: Story and Comic Creator
description: Create a story based on an idea and generate a 4-panel comic strip illustrating the story.

main:
  # Step 1: Generate a story and outline into 4 panels
  - prompt:
      - role: system
        content: You are {{agent.name}}. {{agent.about}}
      - role: user
        content: >
          Based on the idea '{{_.idea}}', write a short story suitable for a 4-panel comic strip.
          Provide the story and a numbered list of 4 brief descriptions for each panel illustrating key moments in the story.
    unwrap: true

  # Step 2: Extract the panel descriptions and story
  - evaluate:
      story: _.split('1. ')[0].trim()
      panels: _.match(/\\d+\\.\\s*(.*?)(?=\\d+\\.\\s*|$)/g)

  # Step 3: Generate images for each panel using the image generator tool
  - foreach:
      in: _.panels
      do:
        tool: image_generator
        arguments:
          description: _

  # Step 4: Generate a catchy title for the story
  - prompt:
      - role: system
        content: You are {{agent.name}}. {{agent.about}}
      - role: user
        content: >
          Based on the story below, generate a catchy title.

          Story: {{outputs[1].story}}
    unwrap: true

  # Step 5: Return the story, the generated images, and the title
  - return:
      title: outputs[3]
      story: outputs[1].story
      comic_panels: outputs[2].map(output => output.image.url)
`;

async function createTask(agent) {
  const task = await client.tasks.create(agent.id, yaml.load(taskYaml));
  return task;
}

Step 3: Execute the Task

async function executeTask(task) {
  const execution = await client.executions.create(task.id, {
    input: { idea: "A cat who learns to fly" }
  });

  // 🎉 Watch as the story and comic panels are generated
  for await (const transition of client.executions.transitions.stream(execution.id)) {
    console.log(transition);
  }

  // 📦 Once the execution is finished, retrieve the results
  const result = await client.executions.get(execution.id);
  return result;
}

Step 4: Chat with the Agent

async function chatWithAgent(agent) {
  const session = await client.sessions.create({ agent_id: agent.id });

  // 💬 Send messages to the agent
  const rl = readline.createInterface({
    input: process.stdin,
    output: process.stdout
  });

  const chat = async () => {
    rl.question("Enter a message (or 'quit' to exit): ", async (message) => {
      if (message.toLowerCase() === 'quit') {
        rl.close();
        return;
      }

      const response = await client.sessions.chat(session.id, { message });
      console.log(response);
      chat();
    });
  };

  chat();
}

// Run the example
async function runExample() {
  const agent = await createAgent();
  const task = await createTask(agent);
  const result = await executeTask(task);
  console.log("Task Result:", result);
  await chatWithAgent(agent);
}

runExample().catch(console.error);

Tip

You can find the full Node.js example here.

Components

Julep is made up of the following components:

  • Julep Platform: The Julep platform is a cloud service that runs your workflows. It includes a language for describing workflows, a server for running those workflows, and an SDK for interacting with the platform.
  • Julep SDKs: Julep SDKs are a set of libraries for building workflows. There are SDKs for Python and JavaScript, with more on the way.
  • Julep API: The Julep API is a RESTful API that you can use to interact with the Julep platform.

Mental Model

Think of Julep as a platform that combines both client-side and server-side components to help you build advanced AI agents. Here's how to visualize it:

  1. Your Application Code:

    • You use the Julep SDK in your application to define agents, tasks, and workflows.
    • The SDK provides functions and classes that make it easy to set up and manage these components.
  2. Julep Backend Service:

    • The SDK communicates with the Julep backend over the network.
    • The backend handles execution of tasks, maintains session state, stores documents, and orchestrates workflows.
  3. Integration with Tools and APIs:

    • Within your workflows, you can integrate external tools and services.
    • The backend facilitates these integrations, so your agents can, for example, perform web searches, access databases, or call third-party APIs.

In simpler terms:

  • Julep is a platform for building stateful AI agents.
  • You use the SDK (like a toolkit) in your code to define what your agents do.
  • The backend service (which you can think of as the engine) runs these definitions, manages state, and handles complexity.

Concepts

Julep is built on several key technical components that work together to create powerful AI workflows:

graph TD
    User[User] ==> Session[Session]
    Session --> Agent[Agent]
    Agent --> Tasks[Tasks]
    Agent --> LLM[Large Language Model]
    Tasks --> Tools[Tools]
    Agent --> Documents[Documents]
    Documents --> VectorDB[Vector Database]
    Tasks --> Executions[Executions]

    classDef client fill:#9ff,stroke:#333,stroke-width:1px;
    class User client;

    classDef core fill:#f9f,stroke:#333,stroke-width:2px;
    class Agent,Tasks,Session core;
Loading
  • Agents: AI-powered entities backed by large language models (LLMs) that execute tasks and interact with users.
  • Users: Entities that interact with agents through sessions.
  • Sessions: Stateful interactions between agents and users, maintaining context across multiple exchanges.
  • Tasks: Multi-step, programmatic workflows that agents can execute, including various types of steps like prompts, tool calls, and conditional logic.
  • Tools: Integrations that extend an agent's capabilities, including user-defined functions, system tools, or third-party API integrations.
  • Documents: Text or data objects associated with agents or users, vectorized and stored for semantic search and retrieval.
  • Executions: Instances of tasks that have been initiated with specific inputs, with their own lifecycle and state machine.

For a more detailed explanation of these concepts and their interactions, please refer to our Concepts Documentation.

Understanding Tasks

Tasks are the core of Julep's workflow system. They allow you to define complex, multi-step AI workflows that your agents can execute. Here's a brief overview of task components:

  • Name and Description: Each task has a unique name and description for easy identification.
  • Main Steps: The core of a task, defining the sequence of actions to be performed.
  • Tools: Optional integrations that extend the capabilities of your agent during task execution.

Types of Workflow Steps

Tasks in Julep can include various types of steps, allowing you to create complex and powerful workflows. Here's an overview of the available step types, organized by category:

Common Steps

  1. Prompt: Send a message to the AI model and receive a response.

    - prompt: "Analyze the following data: {{data}}"
  2. Tool Call: Execute an integrated tool or API.

    - tool: web_search
      arguments:
        query: "Latest AI developments"
  3. Evaluate: Perform calculations or manipulate data.

    - evaluate:
        average_score: "sum(scores) / len(scores)"
  4. Wait for Input: Pause workflow until input is received.

    - wait_for_input:
        info:
          message: "Please provide additional information."
  5. Log: Log a specified value or message.

    - log: "Processing completed for item {{item_id}}"

Key-Value Steps

  1. Get: Retrieve a value from a key-value store.

    - get: "user_preference"
  2. Set: Assign a value to a key in a key-value store.

    - set:
        user_preference: "dark_mode"

Iteration Steps

  1. Foreach: Iterate over a collection and perform steps for each item.

    - foreach:
        in: "data_list"
        do:
          - log: "Processing item {{_}}"
  2. Map-Reduce: Map over a collection and reduce the results.

    - map_reduce:
        over: "numbers"
        map:
          - evaluate:
              squared: "_ ** 2"
        reduce: "sum(results)"
  3. Parallel: Run multiple steps in parallel.

    - parallel:
        - tool: web_search
          arguments:
            query: "AI news"
        - tool: weather_check
          arguments:
            location: "New York"

Conditional Steps

  1. If-Else: Conditional execution of steps.

    - if: "score > 0.8"
      then:
        - log: "High score achieved"
      else:
        - log: "Score needs improvement"
  2. Switch: Execute steps based on multiple conditions.

    - switch:
        - case: "category == 'A'"
          then:
            - log: "Category A processing"
        - case: "category == 'B'"
          then:
            - log: "Category B processing"
        - case: "_"  # Default case
          then:
            - log: "Unknown category"

Other Control Flow

  1. Sleep: Pause the workflow for a specified duration.

    - sleep:
        seconds: 30
  2. Return: Return a value from the workflow.

    - return:
        result: "Task completed successfully"
  3. Yield: Run a subworkflow and await its completion.

    - yield:
        workflow: "data_processing_subflow"
        arguments:
          input_data: "{{raw_data}}"
  4. Error: Handle errors by specifying an error message.

    - error: "Invalid input provided"

Each step type serves a specific purpose in building sophisticated AI workflows. This categorization helps in understanding the various control flows and operations available in Julep tasks.

Advanced Features

Julep offers a range of advanced features to enhance your AI workflows:

Adding Tools to Agents

Extend your agent's capabilities by integrating external tools and APIs:

client.agents.tools.create(
    agent_id=agent.id,
    name="web_search",
    description="Search the web for information.",
    integration={
        "provider": "brave",
        "method": "search",
        "setup": {"api_key": "your_brave_api_key"},
    },
)

Managing Sessions and Users

Julep provides robust session management for persistent interactions:

session = client.sessions.create(
    agent_id=agent.id,
    user_id=user.id,
    context_overflow="adaptive"
)

# Continue conversation in the same session
response = client.sessions.chat(
    session_id=session.id,
    messages=[
      {
        "role": "user",
        "content": "Follow up on the previous conversation."
      }
    ]
)

Document Integration and Search

Easily manage and search through documents for your agents:

# Upload a document
document = client.agents.docs.create(
    title="AI advancements",
    content="AI is changing the world...",
    metadata={"category": "research_paper"}
)

# Search documents
results = client.agents.docs.search(
    text="AI advancements",
    metadata_filter={"category": "research_paper"}
)

For more advanced features and detailed usage, please refer to our Advanced Features Documentation.

Integrations

Julep supports various integrations that extend the capabilities of your AI agents. Here's a list of available integrations and their supported arguments:

Brave Search

setup:
  api_key: string  # The API key for Brave Search

arguments:
  query: string  # The search query for searching with Brave

output:
  result: string  # The result of the Brave Search

BrowserBase

setup:
  api_key: string       # The API key for BrowserBase
  project_id: string    # The project ID for BrowserBase
  session_id: string    # (Optional) The session ID for BrowserBase

arguments:
  urls: list[string]    # The URLs for loading with BrowserBase

output:
  documents: list       # The documents loaded from the URLs

Email

setup:
  host: string      # The host of the email server
  port: integer     # The port of the email server
  user: string      # The username of the email server
  password: string  # The password of the email server

arguments:
  to: string        # The email address to send the email to
  from: string      # The email address to send the email from
  subject: string   # The subject of the email
  body: string      # The body of the email

output:
  success: boolean  # Whether the email was sent successfully

Spider

setup:
  spider_api_key: string  # The API key for Spider

arguments:
  url: string             # The URL for which to fetch data
  mode: string            # The type of crawlers (default: "scrape")
  params: dict            # (Optional) The parameters for the Spider API

output:
  documents: list         # The documents returned from the spider

Weather

setup:
  openweathermap_api_key: string  # The API key for OpenWeatherMap

arguments:
  location: string                # The location for which to fetch weather data

output:
  result: string                  # The weather data for the specified location

Wikipedia

arguments:
  query: string           # The search query string
  load_max_docs: integer  # Maximum number of documents to load (default: 2)

output:
  documents: list         # The documents returned from the Wikipedia search

These integrations can be used within your tasks to extend the capabilities of your AI agents. For more detailed information on how to use these integrations in your workflows, please refer to our Integrations Documentation.

SDK Reference

API Reference

Explore our comprehensive API documentation to learn more about agents, tasks, and executions:

About

API for building multi-step agent workflows. (An orchestrator for AI agents and workflows.)

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 78.4%
  • Jupyter Notebook 20.3%
  • TypeScript 0.4%
  • Dockerfile 0.3%
  • JavaScript 0.3%
  • Shell 0.2%
  • HCL 0.1%