Agentic AI

Agentic AI Architectures

This repository demonstrates multiple Agentic AI patterns and architectures built on top of foundational tools like LangChain, LangGraph, and OpenAI. Specifically, it includes:

• ReAct Agent: Leveraging reasoning and action for dynamic tool use with LangChain.
• PromptChain: Chaining prompts and intermediate steps for complex workflows.
• Agentic RAG: Retrieval-Augmented Generation with agentic control flows.
• Corrective Agentic RAG: Enhanced RAG with document grading and corrective feedback loops.
• Adaptive RAG: An adaptive retrieval pipeline that rewrites queries and invokes web search when needed.
• LangGraph Workflows: Visualizing agent flows and integrating vector stores, web search, and custom tools.

---

Architecture Overview

Agentic AI combines traditional language model prompting with autonomous decision-making and tool usage. This repo explores several patterns:

• ReAct Agent: Integrates chain-of-thought reasoning with action calls to tools (e.g., vectorstore, web search).
• PromptChain: Splits complex tasks into a sequence of specialized prompts with structured intermediate outputs.
• Agentic RAG: Uses a vector retriever to fetch documents, then an agent orchestrates reasoning to form answers.
• Corrective Agentic RAG: Adds a grading step where retrieved documents are evaluated; if none are relevant, a corrective query rewrite triggers web search.
• Adaptive RAG: Extends corrective RAG by dynamically choosing retrieval or web search based on a document grading node; supports fallbacks and query transforms.
• LangGraph: Provides visual representations of these flows, enabling rapid inspection and debugging of agent decision branches.

---

Prerequisites

• Python 3.10+
• OpenAI API Key
• LangChain
• LangGraph (for graph visualizations)
• Other optional dependencies: faiss-cpu or weaviate-client, requests for web search

---

Installation

# Clone this repo
git clone https://github.com/Hitesh-s0lanki/GenAI-AgentAI.git
cd GenAI-AgentAI

# Create a virtual environment
python -m venv venv
source venv/bin/activate  # or `venv\Scripts\activate` on Windows

# Install core dependencies
pip install -r requirements.txt

Set your OpenAI key:

export OPENAI_API_KEY="your_api_key_here"

---

Agent Implementations

ReAct Agent

• Combines reasoning (Thoughts) and actions (e.g., vector store retrieval, API calls).
• Implemented via LangChain's AgentExecutor with Tool definitions.

from langchain.agents import initialize_agent, Tool
from langchain.llms import OpenAI

tools = [
    Tool(name="search", func=web_search, description="Use for online queries."),
    Tool(name="vector_retrieval", func=vector_retrieve, description="Fetch docs.")
]

agent = initialize_agent(tools, OpenAI(temperature=0), agent="react")
response = agent.run("What is the latest news on Agentic AI?")

PromptChain

• Orchestrates a sequence of prompts with structured outputs.
• Useful for multi-step workflows like summarization followed by QA.

from langchain.chains import LLMChain, SequentialChain
from langchain.prompts import PromptTemplate

summary_prompt = PromptTemplate(input_variables=["doc"], template="Summarize: {doc}")
qa_prompt = PromptTemplate(input_variables=["summary","question"], template="Based on the summary, answer: {question}")

chain = SequentialChain(chains=[
    LLMChain(llm=OpenAI(), prompt=summary_prompt),
    LLMChain(llm=OpenAI(), prompt=qa_prompt)
], input_variables=["doc","question"], output_variables=["answer"])

result = chain.run({"doc": text, "question": "..."})

Agentic RAG

• Retrieval-augmented generation pipeline wrapped by an agent.
• Flow: __start__ → agent → vector retrieve → generate → __end__.

Corrective Agentic RAG

• Adds a grade_documents step after retrieval.
• If no documents are relevant, uses transform_query to rewrite and triggers web_search.
• Otherwise proceeds to generate.

Adaptive RAG

• Combines both vector retrieval and web search in a unified graph.
• Dynamically branches based on document relevance and supported content types.

---

LangGraph Visualizations

This repo includes .lg.json graph definitions for each agent, viewable in the LangGraph Playground:

• graphs/agentic_rag.lg.json
• graphs/corrective_rag.lg.json
• graphs/adaptive_rag.lg.json

Open these files to explore interactive flowcharts of decision paths and tool integrations.

---

Usage Examples

1. Agentic RAG Demo

   python examples/agentic_rag_demo.py
   

2. Corrective Agentic RAG

   python examples/corrective_rag_demo.py
   

3. Adaptive RAG with Fallback

   python examples/adaptive_rag_demo.py
   

Results will showcase how each architecture handles missing context, irrelevant documents, and dynamic query rewrites.

---

Contributing

Contributions are welcome! Please open an issue or submit a pull request with:

• New agent patterns or improvements.
• Additional tool integrations (e.g., search engines, custom APIs).
• Bug fixes and performance optimizations.