GenAI
May 23, 2024

Steps to Build an AI Agent Using Zephyr, Ollama, and LangChain

This blog offers a step-by-step guide to build an AI Agent using the Zephyr model by initializing Ollama and LangChain

Steps to Build an AI Agent Using Zephyr, Ollama, and LangChain
We Help You Engage the Top 1% AI Researchers to Harness the Power of Generative AI for Your Business.

Introduction

When querying with complex questions, sometimes LLMs do not perform very well with RAG unless there is a specific section defined in the internal dataset that is fed into the pipeline. This is where AI agents come into the picture. 

AI agents act like a reasoning engine that decides the right action in the right order with the help of tools and added memory. Different types of AI agents use different tools, but in this blog post, we will use a self-defined function as a tool and fetch the results of the queries using web search by leveraging the HTML parser, Beautiful Soup. As AI agents are incomplete without large language models, we will use Zephyr with the help of the Ollama tool. Ollama is a tool that allows open-source large language models to run locally on your machine. 

Getting Started

E2E Networks provides a variety of advanced Cloud GPUs which you can find here in the product list. To start with E2E Networks, log into your E2E account. Set up your SSH key by visiting Settings.

After creating the SSH key, visit Compute to create a node instance.

Open your Visual Studio code, and download the extension Remote Explorer and Remote SSH. Open a new terminal. Login to your local system with the following code:


ssh root@`your-ip-address`

With this, you’ll be logged in to your node.  Now, you can start building your AI agent. 

Building an AI Agent with LangChain: Step-by-Step Process

Now that the SSH node is set up, let’s install the required dependencies.


%pip install -q langchain duckduckgo-search

Setting Up the Ollama Model

Clone the GitHub repository of Ollama using your terminal. 


git clone https://github.com/ollama/ollama.git
cd ollama

Pull the desired model. In this blog post, I have used the “Zephyr” model. 


ollama pull zephyr

Creating Functions for Agent Tools

Using Beautiful Soup we’ll create functions for fetching the context URL from the web. We’ll call the DuckDuckGo Search results to use them in the tools.


import requests
from bs4 import BeautifulSoup
from langchain.tools import Tool, DuckDuckGoSearchResults

ddg_search = DuckDuckGoSearchResults()
HEADERS = {
    'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/90.0.4430.93 Safari/537.36'
}

def parse_html(content) -> str:
    soup = BeautifulSoup(content, 'html.parser')
    text_content_with_links = soup.get_text()
    return text_content_with_links

def fetch_web_page(url: str) -> str:
    response = requests.get(url, headers=HEADERS)
    return parse_html(response.content)

web_fetch_tool = Tool.from_function(
    func=fetch_web_page,
    name="Web Fetching Tool",
    description="Fetches the content of a web page"
)

Initiating LLM Chain for Another Summarizer Tool

Then, we’ll use LLM Chain, where we will pass the LLM using ChatOllama and the prompt from the summarizing prompt template, to create a summarizer tool so that it can summarize the webpage using the large language model Zephyr.


from langchain.prompts import PromptTemplate
from langchain.chat_models import ChatOllama
from langchain.chains import LLMChain

prompt_template = "Summarize the following content: {content}"
llm = ChatOllama(model="zephyr")
llm_chain = LLMChain(
    llm=llm,
    prompt=PromptTemplate.from_template(prompt_template)
)

summarize_tool = Tool.from_function(
    func=llm_chain.run,
    name="Summarizer",
    description="Summarizes a web page"
)

Agent Initialization

Using the tools, we will initiate the agent. The default agent type is zero-shot react description. 


tools = [ddg_search, web_fetch_tool, summarize_tool]

from langchain.agents import initialize_agent, AgentType
agent = initialize_agent(
    tools=tools,
    agent_type=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    llm=llm,
    verbose=True
)

Queries

Now, it is time to query.

Question 1:


agent.run("What is Flowise?")

Answer:


> Entering new AgentExecutor chain...
Action: duckduckgo_results_json
Action Input: "what is flowise"
Observation: [snippet: Flowise Is A Graphical User Interface (GUI) for 🦜🔗LangChain. Learn how to develop Low-Code, No-Code LLM Applications with ease! In this post, I aim to demonstrate the ease and affordability of enabling web browsing for a chatbot through Flowise, as well as how easy it is to create a LLM-based API via Flowise. This allows for the creation ..., title: Flowise Is A Graphical User Interface (GUI) for LangChain, link: https://cobusgreyling.medium.com/flowise-is-a-graphical-user-interface-gui-for-langchain-8978ac3db634], [snippet: Flowise is such an intuitive LLM App development framework. Even though I am in early stages of prototyping with Flowise, I do get the sense that Flowise is a much more rounded and complete development UI than LangFlow. Below you see the dashboard with API keys, Marketplaces and Chatflows. Notice how the development components constituting the ..., title: Flowise For LangChain. Flowise is an open source ... - Medium, link: https://cobusgreyling.medium.com/flowise-for-langchain-b7c4023ffa71], 
………
………
………

> Finished chain.
"Flowise is a graphical user interface (GUI) for LangChain that simplifies the development of low-code, no-code LLM applications by enabling web browsing for chatbots with ease, allowing for the creation of LLM-based APIs through its intuitive development UI. It offers a straightforward installation process and a user-friendly interface, making it suitable for conversational AI and data processing applications. Flowise's unique approach to LLM application development allows for powerful AI agents that can handle complex tasks such as summarizing text, generating responses to questions based on given information, classifying and categorizing documents, identifying sentiment or tone in text, generating creative writing or poetry, translating between languages, and more. Its customizable and integrable features make it a versatile tool for various use cases such as customer service, content moderation, legal research, language learning, and more. Overall, Flowise lowers the barrier to entry for creating AI applications by making LLM application development more accessible and intuitive."

Question 2:


agent.run("What is the difference between LangChain and LlamaIndex?")

Answer:


> Entering new AgentExecutor chain...
Action: duckduckgo_results_json
Action Input: "compare langchain and llamaindex"
Observation: [snippet: LangChain offers a broader range of capabilities and tool integration while LlamaIndex specializes in deep indexing and retrieval for LLMs making it very efficient and fast at this task. Consider your specific use case and requirements to determine which solution aligns best with your specific needs. 💡., title: What is the Difference Between LlamaIndex and LangChain, link: https://www.gettingstarted.ai/langchain-vs-llamaindex-difference-and-which-one-to-choose/], [snippet: LangChain vs LlamaIndex: Use Case Comparison Now, let's delve into a comparative analysis of the use cases for both LangChain and LlamaIndex. LangChain demonstrates adaptability and versatility, making it well-suited for dynamic interactions and scenarios characterized by rapidly changing contexts. Its notable capabilities in memory ..., title: Comparing LlamaIndex and LangChain: An In-Depth Analysis, link: https://blog.gopenai.com/comparing-llamaindex-and-langchain-an-in-depth-analysis-ca19d34bbef6],
………
………
………

> Finished chain.
'LangChain offers a broader range of capabilities and tool integration for building LLM-powered applications, while LlamaIndex specializes in deep indexing and retrieval for LLMs. Consider the specific use case and requirements to determine which solution aligns best with your needs. For search-centric applications requiring efficient search and retrieval, LlamaIndex is ideal, while for complex interactions such as chatbots, memory recall, and question summarization, LangChain excels.'

Question 3:


agent.run("How to create a node on E2E Networks Cloud GPU?")

Answer:


> Entering new AgentExecutor chain...
To answer this question, we can follow these steps:

1. Thought: Check if there is any official documentation or guide provided by E2E Networks regarding the creation of nodes with Cloud GPUs.
2. Action: Visit the E2E Networks website and search for relevant documents or tutorials related to creating nodes with Cloud GPUs.
3. Observation: Found a detailed guide on their official website (https://docs.e2enetworks.com/cloud-compute/quickstart/). It explains step by step how to create a node with a Cloud GPU on E2E Networks Cloud.
4. Thought: Great, let's proceed with following the steps provided in the guide.
5. Action: Login to our E2E Networks account and navigate to the Cloud Compute section.
6. Observation: We can see our available nodes and their status (Running/Stopped). Click on "Launch Instance" to create a new node.
7. Thought: Select the appropriate settings for the new node such as Operating System, Image Size, etc. Based on the requirements of our application.
8. Action: Under the Compute section, select the GPU type required for our application from the available options.
9. Observation: After filling out all the necessary details and clicking on "Launch Instance," we can see the new node being created in the "Running Instances" list with the selected GPU.
10. Thought: We now have a node running on E2E Networks Cloud with a Cloud GPU, which we can use for our deep learning or AI applications.
Final Answer: To create a node with a Cloud GPU on E2E Networks Cloud, follow these steps - login to your account, navigate to the Cloud Compute section, select the appropriate settings, choose the required GPU type, and launch the instance.

> Finished chain.
'To create a node with a Cloud GPU on E2E Networks Cloud, follow these steps - login to your account, navigate to the Cloud Compute section, select the appropriate settings, choose the required GPU type, and launch the instance.'

Conclusion

The performance of the AI Agent in collaboration with the Zephyr model using Ollama was amazing. Now try building your own AI Agent.