Full Agent + Streamlit (LangGraph)

Back to Labs

Production-ready Streamlit application with LangGraph state machines, RAG, moderation layers, guardrails, and explicit control flow.

Why LangGraph?

State Machines

Explicit control flow with nodes and edges

Cycles & Loops

Support for iterative agent behaviors

Conditional Branching

Dynamic routing based on state

Full Observability

Track state at every step of execution

LangChain (AgentExecutor)

Implicit control flow
Tool calling in a loop
Simpler to get started
Less granular control

LangGraph (StateGraph)

Explicit state machine
Custom node transitions
More complex but powerful
Full control over execution

LangGraph Architecture

Step 1: Environment Setup

Recommended: Docker ensures consistent environment and includes all dependencies.

# 1. Make sure Ollama is running
ollama serve
ollama pull llama3.1

# 2. Navigate to the directory
cd "Full Agent + streamlit + langgraph"

# 3. Run with helper script (recommended)
chmod +x scripts/docker-run.sh
./scripts/docker-run.sh run

# OR use docker-compose directly
docker-compose up --build

# 4. Access the app
# Open: http://localhost:8505

Automated: Use the provided PowerShell script to install dependencies.

1. Setup & Environment

# 1. Navigate to the project
cd "Full Agent + streamlit + langgraph"

# 2. Run the automated installer (PowerShell as Admin)
.\setup\windows\install.ps1

# 3. Setup Python environment & dependencies
python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt

2. RAG System (FAISS + Embeddings)

Important: You must initialize the vector database to enable RAG features.

# Generate the FAISS index from PDFs
python src/rag/generate_vector_db.py

Note: If you have an NVIDIA GPU, run .\setup\windows\fix_pytorch_cuda.bat after setup for acceleration.

Automated: Use the shell script to install Homebrew dependencies.

# 1. Installation & Setup
cd "Full Agent + streamlit + langgraph"
chmod +x setup/mac/install.sh && ./setup/mac/install.sh

# 2. Python environment & dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

2. RAG System (FAISS + Embeddings)

Important: You must initialize the vector database to enable RAG features.

# Generate the FAISS index from PDFs
python src/rag/generate_vector_db.py

AI-Assisted: Use GitHub Copilot or any AI coding assistant to automate installation.

How to Use

The project includes a ready-to-use prompt file that you can copy and paste into Copilot, ChatGPT, Claude, or any AI assistant.

Open COPILOT_INSTALL_PROMPT.md in the project root

Copy the content inside the code block

Paste into your AI assistant (Copilot, ChatGPT, etc.)

Follow the AI-generated step-by-step instructions

What the Prompt Includes

System requirements (Python, Ollama, Graphviz)
Key Python packages from requirements.txt
Step-by-step installation for Windows and macOS
Complete project structure reference
Troubleshooting common issues
Quick commands summary table

# Location of the Copilot prompt file:
COPILOT_INSTALL_PROMPT.md

# The file contains a detailed prompt you can use with:
# - GitHub Copilot Chat
# - ChatGPT / GPT-4
# - Claude
# - Any AI coding assistant

# Simply copy the prompt and paste it into your AI tool!

Step 2: RAG System

FAISS + PDF processing + Embeddings

Important: You must initialize the vector database to enable RAG features.

# Generate the FAISS index from PDFs
python src/rag/generate_vector_db.py

Pro Tip: After running the app, you can also go to the "RAG" tab in the interface and click "Rebuild Vector Database" to refresh your knowledge base without using the command line.

Nodes: Functions that process state

Edges: Transitions between nodes

State: Shared data across nodes

Router: Conditional branching

# src/agent/graph.py
from langgraph.graph import StateGraph, END
from typing import TypedDict, Annotated

class AgentState(TypedDict):
    messages: Annotated[list, add_messages]
    tool_calls: list
    final_response: str

def create_agent_graph():
    graph = StateGraph(AgentState)

    # Add nodes
    graph.add_node("agent", agent_node)
    graph.add_node("tools", tool_node)

    # Add edges
    graph.set_entry_point("agent")
    graph.add_conditional_edges(
        "agent",
        router,
        {"tools": "tools", "end": END}
    )
    graph.add_edge("tools", "agent")  # Cycle back

    return graph.compile()

Toxicity: Filters offensive language

Injection: Blocks prompt attacks

PII Masking: Redacts emails/phones

DLP: Prevents data leakage

# src/moderation/manager.py
class ModerationManager:
    def check_input(self, text: str) -> ModerationResult:
        if self._is_toxic(text):
            return ModerationResult(blocked=True, reason="Toxic content")
        if self._is_injection(text):
            return ModerationResult(blocked=True, reason="Injection attempt")
        return ModerationResult(blocked=False, cleaned=self._mask_pii(text))

Chat Tab

Conversational interface with tools

Setup Tab

LLM, RAG, and persistence config

Help Tab

Documentation and examples

Project Structure

Full Agent + streamlit + langgraph/
├── ui.py                    # Main Streamlit interface
├── config/                  # YAML configurations
│   ├── model_config.yaml    # LLM settings
│   └── prompt_templates.yaml
├── src/
│   ├── agent/               # LangGraph Agent
│   │   ├── graph.py         # StateGraph definition
│   │   ├── nodes.py         # Node functions
│   │   └── state.py         # State schema
│   ├── moderation/          # Security Layer
│   │   └── manager.py       # Toxicity, injection, PII
│   ├── guardrails/          # Quality Layer
│   │   └── validators.py    # Citations, tone
│   ├── rag/                 # RAG Implementation
│   │   ├── retriever.py
│   │   └── generate_vector_db.py
│   ├── tools/               # Agent tools
│   └── llm/                 # LLM clients
└── data/
    ├── docs/books/          # PDF documents
    └── embeddings/          # FAISS index

Learning Checklist

I understand the difference between LangChain AgentExecutor and LangGraph StateGraph I can define nodes, edges, and state in a LangGraph I understand how to implement conditional routing with LangGraph I can implement cycles in the agent graph for tool execution I understand the two-layer safety system (Moderation + Guardrails)

Additional Resources

LangGraph Documentation Streamlit Documentation Ollama (Local LLMs) FAISS Documentation

Full Agent + Streamlit (LangGraph)

Why LangGraph?

State Machines

Cycles & Loops

Conditional Branching

Full Observability

LangChain (AgentExecutor)

LangGraph (StateGraph)

LangGraph Architecture

Step 1: Environment Setup

1. Setup & Environment

2. RAG System (FAISS + Embeddings)

2. RAG System (FAISS + Embeddings)

How to Use

What the Prompt Includes

Step 2: RAG System

LangGraph StateGraph

Moderation Layer

Streamlit Interface

Chat Tab

Setup Tab

Help Tab

Project Structure

Learning Checklist

Additional Resources