LangGraph Agent

A conversational AI agent with real-time web search and memory that survives restarts, built as a state graph with LangGraph.

What it does

• Holds conversations with persistent context, checkpointed to SQLite
• Decides on its own when it needs current information and searches the web via Tavily
• Runs on Llama 3.3 70B through Groq — fast inference on a free tier
• Models the agent as an explicit state graph rather than an ad-hoc loop, so routing between "answer" and "search" is inspectable

Stack

Technology	Role
LangGraph	Agent orchestration (state graph)
Groq API (Llama 3.3 70B)	Language model
Tavily	Real-time web search
SQLite	Persistent memory across sessions
Python 3.12 + uv	Language and dependency manager

Architecture

state.py     →  Graph state definition (messages, context)
config.py    →  LLM, tools and memory wiring
graph.py     →  LangGraph graph: nodes and edges
chat.py      →  Interactive conversation loop
main.py      →  Entry point

Installation

git clone https://github.com/albertfsalapi/langgraph-agent
cd langgraph-agent

uv sync
source .venv/bin/activate

cp .env.example .env      # add your API keys

Environment variables

GROQ_API_KEY=
TAVILY_API_KEY=

Both services have a free tier:

• Groq — https://console.groq.com
• Tavily — https://app.tavily.com

Usage

python main.py

Type your message and press Enter. To quit: exit, quit or q.

Conversation history is written to conversations.db automatically and reloaded on the next run.

Roadmap

• More tools: calculator, code executor
• Web interface with Streamlit or Gradio
• Multi-conversation support with dynamic thread_id
• Unit tests for graph nodes