Manual

Reminder: this is in alpha.

Enjoy these sample commands in lieu of a formal doc.

Tiles

Commands in this section are intended for regular users. Most user-facing functionality is available through slash commands in the chat REPL.

CLI Commands


tiles run [MODELFILE_PATH]     # Run a model (uses default if path not provided)
tiles run -r <count>           # Set max relay count for model communication (default: 10)
tiles health                   # Check status of dependencies
tiles memory set-path <path>   # Set the path for memory storage
tiles server start             # Start the daemon server
tiles server stop              # Stop the daemon server

Examples


# Start chatting with the default model
tiles run
 
# Run a specific Modelfile
tiles run ./path/to/Modelfile
 
# Check if all dependencies are properly installed
tiles health
 
# Configure memory storage location
tiles memory set-path ~/.tiles/memory
 
# Start the background server
tiles server start

Tilekit

Commands in this section are meant for developers.

Prompt Optimization

We use DSRs (dspy-rs built into the Tiles CLI to automatically optimize system prompts in Modelfiles.

CLI Commands


tiles optimize <MODELFILE_PATH> [--data <path>] [--model <provider:model-name>]  # Optimize SYSTEM prompt in Modelfile

Examples


# Optimize SYSTEM prompt in a Modelfile
tiles optimize path/to/Modelfile
 
# Optimize with custom training data
tiles optimize ./Modelfile --data training-data.json
 
# Optimize using a specific model
tiles optimize ./Modelfile --model anthropic:claude-3-5-sonnet-20240620

Optimize Command Details

The tiles optimize command refines the SYSTEM prompt in your Modelfile automatically using DSRs (dspy-rs) and the COPRO optimizer.

Basic Usage:


tiles optimize path/to/Modelfile

By default, this will:

Generate 5 synthetic examples automatically if you do not provide your own training data
Use openai:gpt-4o-mini as the optimization model
Update your Modelfile in-place with the improved SYSTEM prompt

Options:

--data <path> - Provide your own training data as a JSON file
--model <provider:model-name> - Specify a different model to use for optimization

Requirements:

Your Modelfile must contain a starting SYSTEM prompt, so the optimizer knows what it should build upon.

Training Data Format:

If supplying your own data, use JSON structured like this:


[
  {
    "input": "User query here",
    "output": "Expected AI response here"
  }
]

Example:

Let’s say you have a Modelfile:


FROM llama3
SYSTEM "You are a helpful assistant."

To optimize its SYSTEM prompt using Claude:


export ANTHROPIC_API_KEY=sk-ant-...
tiles optimize ./Modelfile --model anthropic:claude-3-5-sonnet-20240620

This will rewrite the SYSTEM line with a more effective version based on the optimization process.