kryten-llm 0.2.4

Kryten LLM integration service - provides AI chat responses and interactions

Kryten LLM

AI-powered chat bot service for CyTube, part of the Kryten ecosystem.

Features

• Trigger-Based Responses: Configurable trigger words with probabilities
• Direct Mentions: Responds when mentioned by name
• LLM Integration: Multiple LLM provider support (OpenAI, OpenRouter, local)
• Rate Limiting: Per-user, per-trigger, and global rate limits
• Spam Detection: Automatic spam detection with exponential backoff penalties
• Context Awareness: Tracks chat history and video context
• Hot-Reload: Reload configuration without restart (SIGHUP)
• Service Discovery: Publishes heartbeats and responds to discovery polls
• Dry-Run Mode: Test responses without sending to chat

Requirements

• Python 3.10+
• Poetry
• NATS server
• kryten-py library

Installation

Using Poetry

# Install dependencies
poetry install

# Copy example configuration
cp config.example.json config.json

# Edit configuration with your settings
# See config.example.json for all options

Using pip

pip install kryten-llm

Configuration

Configuration is stored in a JSON file. See config.example.json for a complete example.

Key Sections

Section	Description
nats	NATS connection settings
channels	CyTube channel connections
personality	Bot character and behavior
llm_providers	LLM API configurations
triggers	Trigger words with patterns and probabilities
rate_limits	Rate limiting rules
spam_detection	Spam detection settings
testing	Dry-run and logging options

Environment Variables

Override configuration with environment variables:

export OPENROUTER_API_KEY="your-api-key"
export KRYTEN_LLM_DRY_RUN="true"

Usage

Running the Service

# Using Poetry
poetry run kryten-llm --config config.json

# Direct Python execution
python -m kryten_llm --config config.json

Command Line Options

Option	Description
--config PATH	Path to configuration file (default: config.json)
--log-level LEVEL	Logging level: DEBUG, INFO, WARNING, ERROR
--dry-run	Generate responses but don't send to chat
--validate-config	Validate configuration and exit

Validation Mode

Validate your configuration without starting the service:

poetry run kryten-llm --config config.json --validate-config

Dry-Run Mode

Test responses without sending to chat:

poetry run kryten-llm --config config.json --dry-run

Hot-Reload (POSIX)

Reload configuration without restarting the service:

# Send SIGHUP to reload configuration
kill -HUP $(pgrep -f kryten_llm)

Safe changes that can be hot-reloaded:

• Triggers (patterns, probabilities, enabled status)
• Rate limits
• Spam detection settings
• Personality configuration
• LLM provider settings

Unsafe changes (require restart):

• NATS connection settings
• Channel configuration
• Service name

Production Deployment

Systemd Service

Install the systemd service file:

# Copy service file
sudo cp kryten-llm.service /etc/systemd/system/

# Create config directory
sudo mkdir -p /etc/kryten-llm
sudo cp config.json /etc/kryten-llm/

# Create log directory
sudo mkdir -p /var/log/kryten-llm
sudo chown kryten:kryten /var/log/kryten-llm

# Enable and start service
sudo systemctl daemon-reload
sudo systemctl enable kryten-llm
sudo systemctl start kryten-llm

Service Management

# Check status
sudo systemctl status kryten-llm

# View logs
sudo journalctl -u kryten-llm -f

# Reload configuration
sudo systemctl reload kryten-llm

# Restart service
sudo systemctl restart kryten-llm

Development

Running Tests

# Run all tests
poetry run pytest

# Run with coverage
poetry run pytest --cov=kryten_llm

# Run specific test file
poetry run pytest tests/test_trigger_engine.py -v

Code Quality

# Linting
poetry run ruff check .

# Formatting
poetry run black .

# Type checking
poetry run mypy kryten_llm

Architecture

The service processes messages through a pipeline:

Chat Message
    ↓
MessageListener (parse/filter)
    ↓
TriggerEngine (detect triggers/mentions)
    ↓
RateLimiter (check rate limits)
    ↓
SpamDetector (check spam)
    ↓
ContextManager (gather context)
    ↓
PromptBuilder (build LLM prompt)
    ↓
LLMManager (call LLM API)
    ↓
ResponseValidator (validate response)
    ↓
ResponseFormatter (format for chat)
    ↓
Send to Chat

Service Discovery

The service publishes:

• Startup event: When service starts
• Heartbeat: Every 30s with health status
• Shutdown event: When service stops gracefully

Subscribe to service events:

• kryten.lifecycle.llm.startup
• kryten.heartbeat.llm
• kryten.lifecycle.llm.shutdown

License

MIT License - see LICENSE for details.