GenAI & LLMs | Video 10 | Langchain Agents and Toolkits | Venkat Reddy AI Classes

GenAI & LLMs | Video 10 | Langchain Agents and Toolkits | Venkat Reddy AI Classes

Introduction

Nowadays, in the world of GenAI and large language models (LLMs), one of the most advanced and sought-after topics is Agents. Every company, while solving a problem, is looking to create an agent rather than directly solving the problem.

An agent is a software program that not only aims to get the answer but also validates it to provide the perfect final output. Currently, agents are still in the development phase, but we anticipate having perfect agents in the near future that reliably deliver the results we seek.

Issue with Standalone Large Language Models

Standalone large language models like OpenAI have limitations. For instance, they can't natively perform complex mathematics or search the internet for real-time information. They work using a sequence of words or tokens as input and predict the next sequence. Thus, they struggle with tasks that require real-time data or direct mathematical calculations.

The reasoning capacity of LLMs is good; they understand and generate outputs in human language. However, for tasks beyond their training data, they fall short. Hence, we introduce agents to bridge this gap.

What Are Agents?

Agents are complex software programs connecting LLMs to real-world events, external data, and current information beyond an LLM's training data. They utilize the reasoning capacity of LLMs to get the final answer. Agents can be used as alternatives to implementations we've done so far, such as sequential chains, memory-based tasks, and other complex processes.

An agent determines the sequence of actions needed based on the provided query. It uses a concept called "react" (reasoning and acting) to come up with observations, thoughts, actions, and ultimately, the perfect answer.

Components of Agents

1. Reasoning: The agent interprets the user's query and thinks about the necessary steps.
2. Acting: It utilizes various tools to take actions based on its reasoning.
3. Observation: The feedback or data obtained from the actions taken.

Example Implementation

Here's an example demonstrating how an agent can be constructed to fetch information from Wikipedia and validate it:

from langchain.llms import OpenAI
from langchain.tools import Wikipedia
from langchain.agents import create_agent

llm = OpenAI(api_key="your-api-key")
wiki_tool = Wikipedia()
agent = create_agent(llm, tools=[wiki_tool], agent_type="react")

query = "What is the law related to UPI cyber crimes in India?"
result = agent.run(query)
print(result)

In this example, the agent uses the Wikipedia tool to find information and validate it.

Custom Tools and Toolkits

Agents can also use custom tools. For instance, you can create a custom tool to fetch the current date:

from datetime import datetime
from langchain.tools import Tool

@Tool
def date_tool():
    return datetime.now().strftime("%Y-%m-%d")

llm = OpenAI(api_key="your-api-key")
wiki_tool = Wikipedia()
agent = create_agent(llm, tools=[wiki_tool, date_tool], agent_type="react")

query = "What is today's date?"
result = agent.run(query)
print(result)

Moreover, LangChain offers various toolkits like the CSV toolkit, SQL database toolkit, and Pandas DataFrame toolkit to enhance the capabilities of agents.

Future of Agents

Agents are still evolving. Although currently they may give inconsistent results or hit errors due to limitations like maximum tokens or rate limits, continuous improvements are underway. In the future, agents will be more reliable and could possibly replace numerous human tasks.

The ultimate goal is to have multi-agent systems or organizations of agents working together to provide the final, validated output.

Conclusion

In summary, while agents are highly promising, it's crucial to implement them carefully, considering potential pitfalls like security and consistency. Understand the problem statement before determining whether to use traditional ML models, prompting, fine-tuning an LLM, using a retrieval-augmented generation (RAG) framework, or employing agents.

Keywords

• GenAI
• Large Language Models (LLM)
• Agents
• LangChain
• Wikipedia Tool
• Custom Tools
• SQL Database Toolkit
• Pandas DataFrame Toolkit
• Fine-Tuning
• Reasoning and Acting (React)
• Prompt Ingestion Attack
• Multi-Agent Systems

FAQ

1. What is an agent in the context of Large Language Models (LLMs)? An agent is a complex software program that connects LLMs to real-world data and events, validating and providing perfect final outputs by utilizing various tools and reasoning capabilities.

2. What are the limitations of standalone LLMs? Standalone LLMs cannot perform complex mathematical calculations or fetch real-time data. They rely on pre-existing sequences of words and predictive modeling, which limits their ability to handle current events or numerical computations directly.

3. How can agents overcome the limitations of standalone LLMs? Agents use additional tools to connect LLMs to current and real-world data. They can perform complex calculations, search real-time data, and validate information by reasoning and acting.

4. What is the "react" concept used in agents? "React" stands for Reasoning and Acting. It is a process where the agent interprets a query (reasoning), acts using various tools, makes observations, and continues this cycle until a perfect final answer is derived.

5. Can I create custom tools for my agent? Yes, you can create custom tools for your agent by defining functions and adding decorators to them. These tools can be integrated with pre-existing tools to enhance the agent's capabilities.

6. What are the potential issues with using agents? Agents can sometimes give inconsistent results, face token limitations, or run into security vulnerabilities such as prompt ingestion attacks. It's essential to use them carefully and implement necessary security measures.

7. What types of toolkits are available in LangChain for agents? LangChain offers various toolkits like the CSV toolkit, SQL database toolkit, and Pandas DataFrame toolkit, among others. These toolkits provide specialized functions to enhance the agent's performance.

8. Are agents production-ready? While agents show great potential, they are not yet perfect for production environments due to their current limitations and vulnerabilities. Continuous research and improvements are needed before they become fully reliable.