LangGraph Examples with Ollama

A complete set of examples demonstrating LangGraph with Pydantic and Ollama (local LLM), including advanced pipelines with Human-in-the-Loop (HITL) and a Flask web interface.

🎯 Examples Overview

File	Description	Ollama	HITL	Web UI	Port
simple_graph.py	Basic graph with keyword sentiment	❌	❌	❌	-
llm_graph.py	LLM-powered sentiment analysis	✅	❌	❌	-
chat_loop.py	Multi-turn conversation with loops	✅	❌	❌	-
persistent_chat.py	Conversations with memory/checkpoints	✅	❌	❌	-
tool_agent.py	ReAct agent with tool calling	✅	❌	❌	-
advanced_pipeline.py	Document pipeline with retry/rollback	✅	❌	❌	-
requirements_transformer.py	Requirements doc transformation	✅	✅	❌	-
app.py	Flask Web App for all examples (A-E)	✅	❌	✅	8080
transformer_app.py	Flask Web App for transformation	✅	✅	✅	5001
sample_transformer_app.py	Flask Web App with sample file selector	✅	✅	✅	5002
run_demo.py	Demo Page with agent videos	❌	❌	✅	8000

🚀 Quick Start

🎯 NEW USER? Start here: START_HERE.md - Step-by-step guide to run your first example!

📖 More guides:

• QUICKSTART.md - Get running in 5 minutes
• SETUP.md - Complete step-by-step setup guide with troubleshooting
• COMMANDS.md - All commands in one place
• MODELS.md - Guide to choosing the right Ollama model
• RUN_EXAMPLES.md - How to run each example independently

Prerequisites

Quick Install (Recommended):

# Run the automated installer (after installing Python and Ollama)
# In CMD: install.bat
# In PowerShell: .\install.ps1

Manual Install:

# 1. Install Python dependencies
pip3 install -r requirements.txt

# 2. Install & start Ollama (https://ollama.com/download)
ollama pull llama3.2:1b      # Default model (~1.3GB, very fast, good quality)
# OR for even faster: ollama pull tinyllama (~637MB)
# OR for better quality: ollama pull llama3.2 (~2GB)
ollama serve  # If not running as service (usually runs automatically on macOS)

Note: On macOS, Ollama typically runs as a background service automatically after installation. You can verify it's running by checking ollama list or visiting http://localhost:11434.

Run Examples

# Basic examples (A-E) - use python3 on macOS
python3 simple_graph.py        # A: No LLM
python3 llm_graph.py           # B: LLM sentiment
python3 chat_loop.py --demo    # C: Chat loop (use --demo for non-interactive)
python3 persistent_chat.py     # D: Checkpointing
python3 tool_agent.py          # E: ReAct agent

# Advanced examples
python3 advanced_pipeline.py   # F: Document pipeline
python3 requirements_transformer.py  # G: Transform requirements (CLI)

# Web Apps (run in separate terminals or background)
python3 app.py                 # General examples web UI - Open http://localhost:8080
python3 transformer_app.py     # Transform requirements (Web) - Open http://localhost:5001
python3 sample_transformer_app.py  # Transform with sample selector - Open http://localhost:5002

# Interactive Demo Page
python3 run_demo.py            # Demo page with agent videos - Open http://localhost:8000/demo.html

Note: All examples use python3 command. On some systems, python may point to Python 2.x.

---

📚 Example Details

A. Simple Graph (simple_graph.py)

No LLM required - demonstrates core LangGraph concepts.

START → process → analyze → [positive/negative/neutral] → format → END

Concepts: Pydantic state, nodes, edges, conditional routing

---

B. LLM Sentiment Analysis (llm_graph.py)

Uses Ollama to analyze text sentiment with reasoning.

START → analyze_with_llm → route_by_sentiment → [response nodes] → END

Concepts: OpenAI-compatible API, structured LLM output

---

C. Multi-turn Chat Loop (chat_loop.py)

Interactive conversation that loops until user exits.

START → get_input → process → respond → [continue/exit] → get_input...

Concepts: Graph cycles, conversation history, conditional exit

---

D. Persistent Chat (persistent_chat.py)

Conversations that remember context across sessions.

workflow.invoke(state, config={"configurable": {"thread_id": "user-123"}})

Concepts: MemorySaver checkpointer, thread IDs

---

E. Tool-Calling Agent (tool_agent.py)

ReAct pattern: Reason → Act → Observe.

START → reason → [need_tool/have_answer] → execute_tool → reason...

Tools: get_current_time(), calculate(expr), get_weather(city)

---

F. Advanced Document Pipeline (advanced_pipeline.py)

7-step document processing with retry and rollback on failure.

START → load → parse → extract → prompts → LLM → seed → merge → END
         ↓       ↓                           ↓
      [error_handler] → [retry_router] → retry/rollback

Features:

• Automatic retry (up to 3 attempts per step)
• Rollback to last successful step
• Detailed progress logging
• JSON output with analysis

Run: python advanced_pipeline.py Output: _outputs/pipeline_output.json Diagram: Open advanced_pipeline_diagram.html (also linked in demo page)

---

G. Requirements Transformer (requirements_transformer.py)

Transform generic requirements to IEEE 830 format with Human-in-the-Loop.

START → load_source → load_template → analyze_mapping 
      → [HUMAN REVIEW] → transform → validate 
      → [HUMAN APPROVAL] → generate → END

Features:

• Platform detection (Windows/Linux/macOS, ARM/x86)
• Unique run ID (GUID) for each execution
• LLM-assisted mapping analysis
• Two human review checkpoints

Run: python requirements_transformer.py Output: _outputs/transformed_ieee830_YYYYMMDD_HHMMSS_xxxxxxxx.md

Sample Requirements: See requirements/ folder for sample input files (e-commerce, task manager, blog, weather app, fitness tracker)

---

H. General Examples Web App (app.py)

Interactive web interface for examples A-E (Simple, LLM Sentiment, Chat, Persistent Chat, Tool Agent).

python app.py
# Open http://localhost:8080

Features:

• Tabbed interface for all 5 basic examples
• Real-time Ollama connection status
• Interactive chat interfaces
• Tool agent with live step-by-step reasoning
• No installation needed - runs in browser

Note: Port 8080 is used instead of 5000 to avoid conflicts with macOS AirPlay Receiver.

---

I. Requirements Transformer Web App (transformer_app.py)

Visual web interface for requirements transformation with Human-in-the-Loop.

python transformer_app.py
# Open http://localhost:5001

Features:

• Real-time pipeline progress visualization
• Interactive HITL approve/reject buttons
• Document tabs (Source, Template, Output)
• Platform info and Run ID display
• Live log streaming
• Upload or paste requirements documents

J. Sample Requirements Transformer (sample_transformer_app.py)

Web interface with dropdown selector for pre-written sample requirements files.

python sample_transformer_app.py
# Open http://localhost:5002

Features:

• Dropdown selector for sample requirements files
• Same transformation pipeline as transformer_app.py
• Pre-loaded sample files from requirements/ folder
• Visual progress tracking
• Human-in-the-Loop checkpoints

Sample Files Available:

• E-Commerce Platform
• Task Management Application
• Blog Platform
• Weather Application
• Fitness Tracker

See requirements/README.md for details on available samples.

---

🔧 Configuration

Ollama Settings

All LLM examples use (configurable at top of each file):

OLLAMA_BASE_URL = "http://localhost:11434/v1"
OLLAMA_MODEL = "llama3.2:1b"  # Default: fast 1B model. Change to "tinyllama" for even faster, or "llama3.2" for better quality

Quick Model Switch: To use a faster model, change OLLAMA_MODEL in any Python file, or set environment variable:

# Windows PowerShell
$env:OLLAMA_MODEL="tinyllama"
python llm_graph.py

# macOS/Linux
export OLLAMA_MODEL="tinyllama"
python3 llm_graph.py

Recommended Models (Fastest to Best Quality)

For fastest setup and testing:

Model	Size	Speed	RAM	Best For
tinyllama	~637MB	⚡⚡⚡ Fastest	2GB+	Quick demos, testing
llama3.2:1b	~1.3GB	⚡⚡ Very Fast	2GB+	Fast development
gemma:2b	~1.4GB	⚡⚡ Very Fast	2GB+	Good quality/speed balance

For balanced performance:

Model	Size	Speed	RAM	Best For
llama3.2	~2.0GB	⚡⚡ Balanced	4GB+	Better quality (code uses llama3.2:1b by default)
phi3:mini	~2.3GB	⚡⚡ Balanced	4GB+	Better reasoning tasks

For best quality (slower):

Model	Size	Speed	RAM	Best For
llama3.1:8b	~4.7GB	⚡ Slower	8GB+	Production quality
mistral	~4.1GB	⚡ Slower	8GB+	High quality outputs

💡 Recommendation: Start with tinyllama or llama3.2:1b for fastest setup, then upgrade to llama3.2 if you need better quality.

---

📁 Project Structure

langgraph-learn/
├── simple_graph.py           # A: Basic concepts
├── llm_graph.py              # B: LLM sentiment
├── chat_loop.py              # C: Multi-turn chat
├── persistent_chat.py        # D: Checkpointing
├── tool_agent.py             # E: ReAct agent
├── advanced_pipeline.py      # F: Document pipeline
├── requirements_transformer.py # G: Transform (CLI)
├── app.py                    # H: General examples web UI (port 8080)
├── transformer_app.py        # I: Transform requirements (Web, port 5001)
├── sample_transformer_app.py # J: Transform with sample selector (port 5002)
├── advanced_pipeline_diagram.html # Pipeline visualization
├── requirements.txt          # Dependencies
├── demo.html                 # Interactive demo page with agent videos
├── run_demo.py               # Script to serve demo.html locally
├── requirements/            # Sample requirements documents
│   ├── README.md           # Guide to sample files
│   ├── sample_1_ecommerce.md
│   ├── sample_2_task_manager.md
│   ├── sample_3_blog_platform.md
│   ├── sample_4_weather_app.md
│   └── sample_5_fitness_tracker.md
├── _outputs/                # Generated output files (logs, JSON, MD)
│   └── README.md           # Info about generated files
└── README.md                 # This file

---

🧠 Key Concepts

1. Pydantic State

class WorkflowState(BaseModel):
    messages: list[Message] = Field(default_factory=list)
    current_input: str = ""

2. Node Functions

def process_input(state: WorkflowState) -> dict:
    return {"processed_input": state.user_input.strip()}

3. Conditional Routing

builder.add_conditional_edges(
    "analyze",
    route_function,
    {"option_a": "node_a", "option_b": "node_b"}
)

4. Cycles (Loops)

builder.add_edge("respond", "get_input")  # Creates loop

5. Checkpointing

from langgraph.checkpoint.memory import MemorySaver
checkpointer = MemorySaver()
workflow = builder.compile(checkpointer=checkpointer)

6. Human-in-the-Loop

# Route based on human approval
builder.add_conditional_edges(
    "human_review",
    lambda s: "continue" if s.human_approved else "revise",
    {"continue": "next_step", "revise": "previous_step"}
)

7. Retry/Rollback

# Error handler decides: retry, rollback, or fail
builder.add_conditional_edges(
    "error_handler",
    route_after_error,
    {"step_name": "step_name", "end": END}  # Can route to any step
)

---

🔗 Resources

• LangGraph Documentation
• Ollama
• Pydantic v2
• Flask

---

🖥️ Platform Support

Platform	Architecture	Status
Windows	x86_64	✅ Tested
Windows	ARM64	✅ Tested
Linux	x86_64	✅ Supported
Linux	ARM64	✅ Supported
macOS	Intel	✅ Supported
macOS	Apple Silicon	✅ Supported

Platform is auto-detected and displayed in the web UI and CLI output.

---

🔧 Troubleshooting

Port Conflicts

macOS Port 5000 Issue: macOS uses port 5000 for AirPlay Receiver by default. The app.py web interface uses port 8080 instead to avoid conflicts.

If you need to change ports:

• Edit app.py line 581: app.run(debug=True, host='127.0.0.1', port=8080)
• Edit transformer_app.py to change port 5001 if needed
• Edit sample_transformer_app.py to change port 5002 if needed
• Edit run_demo.py to change port 8000 if needed

Ollama Connection Issues

If examples fail with "Ollama is not running":

1. Check Ollama is installed: ollama --version
2. Start Ollama service: ollama serve (or ensure it's running as a service)
3. Verify model is available: ollama list
4. Pull model if missing: ollama pull llama3.2:1b (or ollama pull llama3.2 for better quality)

Python Version

This project requires Python 3.9+. Check your version:

python3 --version

Dependencies

If you encounter import errors:

pip3 install -r requirements.txt

For ARM Macs (Apple Silicon), all dependencies should install without issues.

---

📄 Output Files

When running examples, the following files may be generated:

• _outputs/pipeline_output.json - Output from advanced_pipeline.py with document analysis
• _outputs/pipeline_run.log - Execution log from pipeline runs
• _outputs/app.log - Logs from running app.py
• _outputs/transformer_app.log - Logs from running transformer_app.py
• _outputs/transformed_ieee830_YYYYMMDD_HHMMSS_xxxxxxxx.md - Generated requirements documents from transformer
• _outputs/output_*.md - Various output files from example runs
• advanced_pipeline_diagram.html - Visual diagram of the pipeline (open in browser, also linked in demo page)

Note: All generated output files are stored in the _outputs/ directory. These files are automatically created when running examples and can be safely deleted. See _outputs/README.md for more details.

✅ Testing Status

All examples have been tested and verified working on:

• ✅ macOS (Apple Silicon) - All examples passing
• ✅ Python 3.9.6
• ✅ Ollama with llama3.2:1b model (default)
• ✅ All web interfaces accessible
• ✅ All syntax errors fixed in app.py

🎬 Interactive Demo Page

View all examples with AI agent explanation videos:

python run_demo.py

Then open http://localhost:8000/demo.html in your browser. The demo page includes:

• Interactive example selector
• AI agent video explanations for each example
• Code overviews with detailed snippets
• "Try It Out" instructions for running each example independently

📝 Recent Updates

• Sample Requirements Transformer: Added sample_transformer_app.py with dropdown selector for sample files (port 5002)
• Sample Requirements Folder: Created requirements/ folder with 5 sample requirements documents
• Interactive Demo: Added demo.html with agent videos and detailed instructions
• Code Overviews: Enhanced code examples in demo page for all A-I examples
• Pipeline Diagram Link: Added link to advanced_pipeline_diagram.html in demo page for Advanced Pipeline example
• Independent Execution: All examples can run independently (see RUN_EXAMPLES.md)
• Model Default: Updated to llama3.2:1b as default for faster performance
• Compressed Avatars: All videos now use compressed versions from avatars/compressed_avatars/ for faster loading
• Port Fix: app.py uses port 8080 instead of 5000 to avoid macOS AirPlay Receiver conflicts
• Syntax Fixes: Fixed indentation errors in chat_loop.py and advanced_pipeline.py
• Dark Theme: Updated sample_transformer_app.py with dark theme for better contrast
• Verified: All examples tested and working