Incorporating human-in-the-loop in agentic logic via LangGraph

Prerequisites

To run this notebook, you need to follow the steps from here and generate an API key from NVIDIA API Catalog.

Please ensure you have the following dependencies installed :

• langchain
• jupyterlab==4.0.8
• langchain-core
• langchain-nvidia-ai-endpoints==0.2.0
• markdown
• colorama

you will also need to install the following -

This notebook will walk you though how to incoporate human-in-the-loop into a multi-agents pipeline in a minimalistic examples.

The cognitive agentic architecture will look like the below :

We will first construct the 2 agents in the middle :

• Using meta/llama-3.1-405b-instruct to construct the 2 agents, each will be created with LCEL expression

• then we will give each agent one tool to use to achieve the task

The task at hand is creating promotion assets with text and image for social medial promotion. We are aiming for something similar to the below ...

Just like in real world, a human in charge of the task will delegate tasks to specalist writer to writ the promotion text and assign a digital artist for the artworks.

In this scenario, we will let human assign an agent ( either ContentCreator or DigitalArtist ) just like the flow depicted above.

Note: As one can see, since we are using NVIDIA AI Catalog as an API, there is no further requirement in the prerequisites about GPUs as compute hardware

We will prepare the 2 agents , each is made out of LCEL expression

For simplicity , each agent will be given one tool to use.

• a content_creator agent which will create promotion message per input product_desc
• an digital_artist agent what is able to create visually appealing image from the promotion title

Step 1 : construct content_creator agent

in order to construct the content_creator agent we need the following :

• system prompt which anchor the task for the agent

• provide a seeded product desc

• a powerful LLM llama3.1-405b from NVIDIA NIM

• using with_structured_output for formatting

Step 2 : we will now create digital_artist agent

We will equip the digital_artist with the following :

• a text-to-image model stableXL-turbo from NVIDIA NIM
• wrap this tool into llm with llm.bind_tools
• construct our digital_artist agent with LCEL expression

a text-to-image model stableXL-turbo from NVIDIA NIM

Wrap the tool into llm with llm.bind_tools

creating digital_artist using LCEL chain

Step 3 - Embed Human-in-the-loop agentic logic with LangGraph

• construct a get_human_input function to integrate into the first node of LangGraph putting Human-in-the-loop deciding which tool to use
• establish State to keep track of the internal states
• create functions as graph nodes for LangGraph
• compose the agentic cognitive logic in langGraph by connecting the nodes and edges

construct a get_human_input function to integrate into the first node of LangGraph

putting Human-in-the-loop deciding which tool to use

establish State to keep track of the internal states

create functions as graph nodes for LangGraph

compose the agentic cognitive logic in langGraph by connecting the nodes and edges

time to test this out

now we will use the output from the ContentCreator agent to go for a 2nd round to generate beautiful image for this promotion

let's try to print this out using markdown