Notebooks
N
NVIDIA Intro Agents

Intro Agents

gpu-accelerationretrieval-augmented-generationllm-inferencetensorrtnvidia-generative-ai-exampleslarge-language-modelsmicroservicetriton-inference-servercommunityLLMagentic-llmragnemoautonomous_5g_slicing_lab
alph-notebooks/nvidia-generative-ai-examples / intro_agents.ipynb

Nvidia Logo

Introduction to LLM Agents in LangGraph and LangChain

This notebook provides an introduction to building LLM-based agents using LangGraph and LangChain. It covers the fundamental implementations required for developing agentic workflows. By the end of this notebook, you will have an understanding of the following concepts:

  1. NVIDIA NIM Endpoints – Learn how to integrate and use NVIDIA NIM endpoints for efficient inference.
  2. Tool Calling using bind_tools – Understand how to define and bind tools within an agent.
  3. Agents using create_react_agent – Explore the creation of ReAct (Reasoning + Acting) agents using Langgraph.

This notebook serves as a introduction to implementing intelligent agents with modular and scalable workflows.

Set up

[3]

Accessing Nvidia NIM endpoints

Here is a list of available models, we will be using meta/llama-3.1-70b-instruct for this notebook.

[ ]

We'll set up the LLM through LangChain's ChatNVIDIA functionality, which provides an interface to NVIDIA NIM chat models. It offers connection to both hosted and local NIMs (Mistral, Llama, etc.), tool calling capabilities, streaming functionality, etc. Here is an example on how to implement that

[5]
[6]
I'm just a computer program, so I don't have feelings, but thanks for asking! How can I assist you today?

Tool calling with .bind_tools()

In LangChain, tool calling allows LLMs to invoke external functions, APIs, or utilities dynamically, extending LLM capabilities beyond text generation. We define tools with the @tool decorator, and bind them to llm with .bind_tool() function. That tells the LLM which tools are available for using. LLM calls these functions with the proper arguments depending on the prompt it receives.

[7]
/home/nvidia/.local/lib/python3.10/site-packages/langchain_nvidia_ai_endpoints/chat_models.py:591: UserWarning: Model 'nvdev/meta/llama-3.1-70b-instruct' is not known to support tools. Your tool binding may fail at inference time.
  warnings.warn(

LLM intelligently decides where to call a tool or not depending on the prompt.

[8]
AIMessage(content="Hello! It's nice to meet you. Is there something I can help you with or would you like to chat?", additional_kwargs={}, response_metadata={'role': 'assistant', 'content': "Hello! It's nice to meet you. Is there something I can help you with or would you like to chat?", 'token_usage': {'prompt_tokens': 283, 'total_tokens': 307, 'completion_tokens': 24}, 'finish_reason': 'stop', 'model_name': 'nvdev/meta/llama-3.1-70b-instruct'}, id='run-34a9624d-e628-4af5-a5c4-f3ceafc8533a-0', usage_metadata={'input_tokens': 283, 'output_tokens': 24, 'total_tokens': 307}, role='assistant')
[9]
[{'name': 'multiply',
,  'args': {'a': 56, 'b': 64},
,  'id': 'chatcmpl-tool-fb0629aaebde42b4ba91e4d36c93a431',
,  'type': 'tool_call'}]

Using .bind_tools(), LLM tells us which tool to call, with which arguments. But it doesn't actually invoke the tool. To invoke the tool intelligently we use the .create_react_agent() prebuilt function from Langgraph.

Prebuilt ReAcT agent using .create_react_agent()

ReAct (Reasoning + Acting) is an agent architecture, based on this paper ReAct: Synergizing Reasoning and Acting in Language Models, that combines step-by-step reasoning with tool use. LangGraph provides a prebuilt function create_react_agent to easily implement this architecture. It uses .bind_tools() under the hood to attach tools to the language model.

[48]
/home/nvidia/.local/lib/python3.10/site-packages/langchain_nvidia_ai_endpoints/chat_models.py:591: UserWarning: Model 'nvdev/meta/llama-3.1-405b-instruct' is not known to support tools. Your tool binding may fail at inference time.
  warnings.warn(
[49]
'The answer is 50.'
[55]
================================ Human Message =================================

What is (5 squared) multiplied by 2?
================================== Ai Message ==================================
Tool Calls:
  square (chatcmpl-tool-3fbc1698aad645b7a9b1a38ebfc3c1f4)
 Call ID: chatcmpl-tool-3fbc1698aad645b7a9b1a38ebfc3c1f4
  Args:
    n: 5
================================= Tool Message =================================
Name: square

25
================================== Ai Message ==================================
Tool Calls:
  multiply (chatcmpl-tool-557233293c6c4011a6d54fcfd2633d88)
 Call ID: chatcmpl-tool-557233293c6c4011a6d54fcfd2633d88
  Args:
    a: 25
    b: 2
================================= Tool Message =================================
Name: multiply

50
================================== Ai Message ==================================

The answer is 50.

Now that you are familiar with the basics of building agents, we will explore how to build a 5G network reconfiguration agent. Please refer to agentic_pipeline-DLI.ipynb for more details.

[ ]
[5]
/home/nvidia/.local/lib/python3.10/site-packages/langchain_nvidia_ai_endpoints/chat_models.py:591: UserWarning: Model 'nvdev/meta/llama-3.1-70b-instruct' is not known to support tools. Your tool binding may fail at inference time.
  warnings.warn(
[6]
/home/nvidia/.local/lib/python3.10/site-packages/langchain_nvidia_ai_endpoints/chat_models.py:814: UserWarning: Model 'nvdev/meta/llama-3.1-70b-instruct' is not known to support structured output. Your output may fail at inference time.
  warnings.warn(
[7]
WeatherResponse(conditions='Cloudy')
[ ]