LangGraph Practice

This repository is a hands-on learning project for LangGraph, LangChain, and Gemini-powered agents. It starts with small graph concepts, then builds up to practical agents that can chat, remember conversation history, call tools, and retrieve information from a PDF.

What This Repo Covers

• Defining graph state with TypedDict
• Creating LangGraph nodes as normal Python functions
• Connecting nodes with normal edges
• Routing execution with conditional edges
• Building loops inside a graph
• Passing multiple values through state
• Calling Google Gemini models through langchain-google-genai
• Keeping manual conversation memory
• Binding tools to an LLM
• Building a ReAct-style tool-using agent
• Building a RAG agent with PDF loading, chunking, embeddings, Chroma, retrieval, and tool calls

Project Structure

.
├── Agent1/
│   └── simple_llm_bot.py
├── Agent2/
│   └── manual_memory_agent.py
├── Agent3/
│   └── ReAct_agent.py
├── Agent4/
│   ├── RAG.py
│   └── Stock_Market_Performance_2024.pdf
├── LangGraph_concepts/
│   ├── hello_world_agent.py
│   ├── processor_multiple_inputs.py
│   ├── mutliple_nodes_sequential_agent.py
│   ├── operation_selector.py
│   ├── conditional_edges.py
│   └── looping_graph.py
├── pyproject.toml
└── README.md

Concept Examples

LangGraph_concepts/hello_world_agent.py introduces the smallest possible graph: one state type, one node, one compiled graph, and one invocation.

LangGraph_concepts/processor_multiple_inputs.py shows how a graph state can carry multiple inputs, such as a name and a list of numbers.

LangGraph_concepts/mutliple_nodes_sequential_agent.py demonstrates sequential node execution. Each node updates the same state and passes it to the next node.

LangGraph_concepts/operation_selector.py shows how one node can choose behavior based on a value in state.

LangGraph_concepts/conditional_edges.py demonstrates conditional routing between different math nodes.

LangGraph_concepts/looping_graph.py demonstrates a graph loop with a guessing game that keeps running until a condition is met.

Agents Built

Agent1/simple_llm_bot.py is a simple Gemini chat bot. It wraps a single LLM call inside a LangGraph node and runs in a terminal loop until the user types exit.

Agent2/manual_memory_agent.py extends the simple bot with manual memory. It stores HumanMessage and AIMessage objects in a Python list and passes the full conversation back into the graph on every turn.

Agent3/ReAct_agent.py is a ReAct-style tool agent. The LLM can request math tools, LangGraph routes to a ToolNode, and the model receives tool results before producing the final answer.

Agent4/RAG.py is a retrieval-augmented generation agent. It loads a stock market PDF, splits it into chunks, embeds those chunks with Google Gemini embeddings, stores them in Chroma, exposes retrieval as a tool, and lets the LLM answer questions using retrieved document context.

Setup

Install dependencies with uv:

uv sync

Create a .env file in the repo root:

GEMINI_API_KEY=your_google_gemini_api_key

The .env file should stay local and should not be committed.

Running Examples

Run any script with uv run python:

UV_CACHE_DIR=/tmp/uv-cache uv run python LangGraph_concepts/hello_world_agent.py
UV_CACHE_DIR=/tmp/uv-cache uv run python Agent1/simple_llm_bot.py
UV_CACHE_DIR=/tmp/uv-cache uv run python Agent3/ReAct_agent.py
UV_CACHE_DIR=/tmp/uv-cache uv run python Agent4/RAG.py

The UV_CACHE_DIR=/tmp/uv-cache prefix is useful in restricted environments where the default uv cache directory is not writable.

Notes

• The Gemini examples require a valid API key with access to the configured model.
• If you see RESOURCE_EXHAUSTED, your key or project has hit a quota limit.
• If you see PERMISSION_DENIED, check whether the API key is valid, leaked, restricted, or missing access to the selected model.
• The RAG example needs Agent4/Stock_Market_Performance_2024.pdf to exist.