LangChain Onboarding Project

Purpose

This repository is for onboarding SWE interns to LangChain by building a simple AI agent. Each engineer will learn how to use LangChain’s core components — models, tools, and agents — to create a runnable prototype demonstrating structured reasoning and output generation.

---

Setup Instructions

1. Clone the Repository

git clone https://github.com/saanviibrahim45/langchain-practice.git
cd langchain-practice

2. Create and Activate a Virtual Environment

python3 -m venv .venv
source .venv/bin/activate   # macOS/Linux
.venv\Scripts\activate      # Windows

3. Install Dependencies

pip install -r requirements.txt

4. Verify Installation and API Key

Check that all key libraries are installed and importable:

pip show langchain langgraph langchain-openai python-dotenv

Then confirm imports work and your environment is set up correctly:

python - <<'PY'
try:
    import langgraph, langchain, langchain_openai, dotenv
    print("Imports OK")
except Exception as e:
    print("Import error:", e)
PY

You should see: Imports OK

Next, verify your OpenAI key is loaded:

python - <<'PY'
import os
print("OPENAI_API_KEY present:", bool(os.environ.get("OPENAI_API_KEY")))
PY

You should see: OPENAI_API_KEY present: True

---

5. Test Model Connection

Confirm that your OpenAI model can respond:

python - <<'PY'
from langchain_openai import ChatOpenAI
m = ChatOpenAI(model="gpt-4o-mini", temperature=0)
resp = m.invoke([{"role":"user","content":"Say hello in one short sentence."}])
print(resp.content)
PY

Expected output: Hello there! How can I help you today?

---

6. Run the Example Agent

From the project root, run:

python src/base_agent.py

You should see:

LangGraph base agent ready. Type 'exit' to quit.
Use 'calc: 2+2' to try the calculator tool.

Example Run

You: calc: 5+5
AI: The result is 10.
You: What did I just calculate?
AI: You just calculated 5 plus 5, which equals 10.

---

Deliverable Expectations

Each SWE must:

1. Work in their own branch (named after them) and add a new file in src/ to build their agent.
2. Implement one working LangChain agent with structured output (JSON or clearly formatted text).
3. Write a short README.md in their folder explaining what their agent does and how it works.
4. Ensure their .env file is configured so their code runs locally end-to-end.

When finished, submit a pull request or share your branch for review.

---

Notes

• Keep your API keys private — never commit .env files.
• If you add new dependencies, update requirements.txt.
• Write clean, modular, and well-commented code.
• Test your agent manually using interactive input.

---

Example Command Summary

git clone https://github.com/saanviibrahim45/langchain-practice.git
cd langchain-practice
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
export OPENAI_API_KEY="your_key_here"
python src/base_agent.py

---

Technical Walkthrough: Base LangGraph Agent

The included src/base_agent.py demonstrates how to build a minimal LangGraph-powered agent using LangChain’s OpenAI integration. It models an LLM-driven reasoning flow using modular nodes that communicate through a shared state.

Core Concepts

1. StateGraph and MessagesState

• StateGraph defines a directed workflow of agent steps.
• Each node processes and returns a MessagesState (a structured list of HumanMessage, AIMessage, and SystemMessage objects).
• This enables memory and context passing across nodes.

2. Nodes and Edges

• Nodes are Python functions that take in state and return updates.
• Edges define the flow of execution (START → tool_node → llm_node → END).

Node	Purpose
tool_node	Detects calculator commands like calc: 2+2 and appends tool output
llm_node	Reads recent messages and produces an AI response via OpenAI

3. Tools

• A “tool” is a callable function the agent can invoke when needed.

• The calculator function (calculator_tool) is a simple example — future tools can include:

  • Web search (e.g. Tavily, Serper)
  • Data analysis
  • File summarization

4. Execution Loop

• The interactive loop lets you type messages directly into the terminal.
• Each message is wrapped as a HumanMessage and passed through the compiled graph.
• The agent then decides how to respond based on the state.

---

Example Run

You: calc: 10*(5+2)
AI: The result is 70.
You: What did I just calculate?
AI: You just calculated 10 times (5 plus 2), which equals 70.