Skip to content

06-models.md

Latest commit

 

History

History
482 lines (348 loc) · 12.5 KB

06-models.md

File metadata and controls

482 lines (348 loc) · 12.5 KB
layout default
title Aider Tutorial - Chapter 6: Model Configuration
nav_order 6
has_children false
parent Aider Tutorial

Chapter 6: Model Configuration

Welcome to Chapter 6: Model Configuration. In this part of Aider Tutorial: AI Pair Programming in Your Terminal, you will build an intuitive mental model first, then move into concrete implementation details and practical production tradeoffs.

Configure and optimize different AI models for various coding tasks, including cost optimization, performance tuning, and model selection strategies.

Overview

Different AI models have different strengths, costs, and performance characteristics. This chapter covers how to configure and optimize Aider for various models and use cases.

Model Selection Strategy

Understanding Model Characteristics

Model Strengths Best For Cost
Claude 3.5 Sonnet Best overall performance, complex reasoning Architecture design, refactoring, debugging High
GPT-4o Fast, good all-rounder, multimodal General coding, quick iterations Medium-High
GPT-4o Mini Cost-effective, fast Simple tasks, prototyping Low
Claude 3 Haiku Fast, good for coding Routine tasks, documentation Low-Medium
Gemini 1.5 Pro Long context, multimodal Large codebases, documentation Medium
Local Models Privacy, offline Sensitive code, no API costs Free

Task-Based Model Selection

# Architecture and design - use most capable model
aider --model claude-3-5-sonnet-20241022

# Complex refactoring - needs strong reasoning
aider --model claude-3-5-sonnet-20241022

# Routine coding tasks - use faster/cheaper model
aider --model gpt-4o-mini

# Documentation and comments - can use smaller model
aider --model claude-3-haiku-20240307

# Large codebase work - use long context model
aider --model gemini/gemini-1.5-pro

Architect Mode

When to Use Architect Mode

Architect mode uses two models: a powerful "architect" model for planning and a fast "editor" model for implementation.

# Use for complex multi-file changes
aider --architect \
      --model claude-3-5-sonnet-20241022 \
      --editor-model gpt-4o-mini

# Benefits:
# - Claude does complex planning and reasoning
# - GPT-4o-mini does fast, accurate editing
# - Cost-effective for complex tasks

Architect Mode Workflow

# Architect model analyzes the request and plans changes
> Refactor the entire authentication system to use dependency injection

# Architect creates detailed plan:
# 1. Extract interfaces for authentication services
# 2. Create dependency injection container
# 3. Refactor existing classes to use DI
# 4. Update initialization code
# 5. Add configuration for DI container

# Editor model implements each step quickly and accurately

Cost Optimization

Model Cost Comparison

# Cost per 1K tokens (approximate, as of 2024)
# Input tokens:
# - Claude 3.5 Sonnet: $3/1K
# - GPT-4o: $2.50/1K
# - GPT-4o Mini: $0.15/1K
# - Claude 3 Haiku: $0.25/1K

# Output tokens:
# - Claude 3.5 Sonnet: $15/1K
# - GPT-4o: $10/1K
# - GPT-4o Mini: $0.60/1K
# - Claude 3 Haiku: $1.25/1K

Cost Optimization Strategies

# Use smaller models for routine tasks
aider --model gpt-4o-mini

# Use architect mode for complex tasks
aider --architect --model claude-3-5-sonnet-20241022 --editor-model gpt-4o-mini

# Be specific in prompts to reduce back-and-forth
> Add input validation to the user registration form with specific rules: email format, password strength (8+ chars, uppercase, lowercase, number), username uniqueness

# Use /diff to review before accepting expensive changes
> /diff

Monitoring Usage

# Aider shows token usage
Aider v0.50.0
Models: claude-3-5-sonnet-20241022 with diff edit format
Git repo: .git with 12 files
Repo-map: using 1024 tokens  # Input tokens for context
API calls: 5 total, 1200 input tokens, 800 output tokens  # Running totals

Performance Tuning

Token Limits and Context Management

# Adjust repo-map tokens for your codebase size
export AIDER_MAP_TOKENS="2048"  # Increase for larger codebases

# Maximum chat history tokens
export AIDER_MAX_CHAT_HISTORY_TOKENS="8192"

# Configuration file
cat > .aider.conf.yml << EOF
map-tokens: 4096
max-chat-history-tokens: 16384
EOF

Model-Specific Optimizations

# Claude models work better with explicit instructions
> Implement a REST API using Flask with the following requirements:
> - Use blueprints for organization
> - Include proper error handling
> - Add input validation with marshmallow
> - Use SQLAlchemy for database operations

# GPT models respond well to examples
> Create a function similar to the existing validate_email function but for phone numbers. Use regex pattern ^\+?1?\d{9,15}$

# Local models may need simpler instructions
ollama serve  # Start Ollama server first
aider --model ollama/llama3.1:70b

Model Configuration Files

Project-Specific Configuration

# .aider.conf.yml - Project-specific settings
model: claude-3-5-sonnet-20241022
editor-model: gpt-4o-mini
auto-commits: true
dark-mode: true
map-tokens: 2048
max-chat-history-tokens: 4096

# File type associations (for better context)
file-associations:
  - "*.py": "python"
  - "*.js": "javascript"
  - "*.ts": "typescript"

User-Specific Configuration

# ~/.aider.conf.yml - Global user preferences
model: gpt-4o
auto-commits: true
dark-mode: true
editor: code
git: true
gitignore: true

# API keys (or use environment variables)
# openai-api-key: sk-...
# anthropic-api-key: sk-ant-...

Environment-Specific Models

# Development - use faster models
export AIDER_MODEL="gpt-4o-mini"

# Production code - use most capable models
export AIDER_MODEL="claude-3-5-sonnet-20241022"

# Documentation - use cost-effective models
export AIDER_MODEL="claude-3-haiku-20240307"

Working with Local Models

Ollama Integration

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull models
ollama pull llama3.1:70b
ollama pull mistral:7b
ollama pull codellama:34b

# Use with Aider
aider --model ollama/llama3.1:70b
aider --model ollama/mistral:7b
aider --model ollama/codellama:34b

Local Model Advantages

# Privacy - no data sent to external APIs
aider --model ollama/llama3.1:70b

# Cost - no API charges
aider --model ollama/mistral:7b

# Offline capability
aider --model ollama/codellama:34b

Local Model Limitations

# Generally slower than cloud models
# May have less up-to-date knowledge
# Smaller context windows
# May need simpler prompts

Cloud Model Providers

OpenAI Models

# GPT-4o (recommended for most users)
aider --model gpt-4o

# GPT-4o Mini (cost-effective)
aider --model gpt-4o-mini

# GPT-4 Turbo (good balance)
aider --model gpt-4-turbo

# Azure OpenAI
export AZURE_OPENAI_API_KEY="your-key"
export AZURE_OPENAI_ENDPOINT="https://your-resource.openai.azure.com/"
export AZURE_OPENAI_DEPLOYMENT="gpt-4"
aider --model azure/gpt-4

Anthropic Models

# Claude 3.5 Sonnet (best overall)
aider --model claude-3-5-sonnet-20241022

# Claude 3 Opus (most capable, expensive)
aider --model claude-3-opus-20240229

# Claude 3 Haiku (fast, cost-effective)
aider --model claude-3-haiku-20240307

# Claude 3 Sonnet (good balance)
aider --model claude-3-sonnet-20240229

Google Models

# Gemini 1.5 Pro (long context)
aider --model gemini/gemini-1.5-pro

# Gemini 1.0 Pro (faster, cheaper)
aider --model gemini/gemini-1.0-pro

Model Switching During Sessions

Dynamic Model Switching

# Start with one model
aider --model gpt-4o-mini

# Switch models mid-session (if needed)
> /model claude-3-5-sonnet-20241022

# Aider will switch models for subsequent requests

Task-Based Switching

# Use different models for different phases
# Planning phase - use capable model
aider --model claude-3-5-sonnet-20241022

# Implementation phase - switch to faster model
> /model gpt-4o-mini

# Continue with faster model for routine tasks

Troubleshooting Model Issues

API Rate Limits

# If you hit rate limits, switch to different model or wait
# OpenAI: 10,000 RPM for GPT-4, higher for GPT-3.5
# Anthropic: 50 requests per minute

# Use retry logic (Aider handles this automatically)
# Or switch to a different provider
aider --model claude-3-haiku-20240307

Model-Specific Errors

# Claude context length exceeded
# Solution: Reduce repo-map tokens or be more specific about files
export AIDER_MAP_TOKENS="1024"

# GPT model hallucinations
# Solution: Be more specific in prompts and review changes carefully
> /diff

# Local model slow responses
# Solution: Use smaller local models or cloud models for speed
aider --model ollama/mistral:7b

Cost Monitoring

# Monitor API usage and costs
# OpenAI: https://platform.openai.com/usage
# Anthropic: https://console.anthropic.com/

# Set spending limits
# Use cost-effective models for routine tasks
aider --model gpt-4o-mini

Advanced Model Features

Multimodal Models

# GPT-4o and Gemini support images
# Can analyze screenshots, diagrams, etc.
> Look at this architecture diagram [attach image] and implement the user service according to this design

# Useful for:
# - UI implementation from designs
# - Code generation from diagrams
# - Documentation from screenshots

Function Calling

# Some models support function calling
# Aider can use this for external integrations
> Run the test suite and fix any failing tests

# Model can call testing functions directly

Best Practices for Model Selection

Match Model to Task Complexity

# Simple tasks: GPT-4o Mini or Claude Haiku
> Add type hints to this function

# Medium complexity: GPT-4o or Claude Sonnet
> Create a REST API endpoint with validation

# High complexity: Claude Opus or GPT-4
> Design and implement a microservices architecture

# Very large tasks: Use architect mode
aider --architect --model claude-3-5-sonnet-20241022 --editor-model gpt-4o-mini

Cost-Performance Balance

# Budget-conscious development
# Use GPT-4o Mini for most tasks (85% cost savings vs GPT-4)
aider --model gpt-4o-mini

# Reserve expensive models for critical tasks
aider --model claude-3-5-sonnet-20241022  # Only when needed

Team Model Standards

# .aider.conf.yml for team consistency
model: gpt-4o-mini          # Default for all developers
architect: claude-3-5-sonnet-20241022  # For complex tasks
editor-model: gpt-4o-mini   # Fast implementation

# Override for specific projects
# high-security-project: use local models
# performance-critical: use Claude Opus

Summary

In this chapter, we've covered:

  • Model Selection: Choosing the right model for different tasks
  • Architect Mode: Using two models for planning and implementation
  • Cost Optimization: Balancing cost and performance
  • Performance Tuning: Token limits and context management
  • Configuration: Project and user-specific settings
  • Local Models: Privacy and cost benefits of local models
  • Cloud Providers: OpenAI, Anthropic, and Google model options
  • Troubleshooting: Handling rate limits and errors
  • Advanced Features: Multimodal and function calling capabilities

Key Takeaways

  1. Task Matching: Choose models based on task complexity and requirements
  2. Cost Awareness: Use cost-effective models for routine tasks
  3. Architect Mode: Leverage two-model approach for complex work
  4. Local Options: Consider privacy and offline capabilities
  5. Configuration: Set up project-specific model preferences
  6. Performance Tuning: Adjust token limits for your codebase
  7. Troubleshooting: Handle API limits and model-specific issues

Next Steps

Now that you can configure models effectively, let's explore voice workflows and automation features.

Ready for Chapter 7? Voice & Workflows

Generated for Awesome Code Docs

Depth Expansion Playbook

Source Code Walkthrough

aider/models.py

Model configuration and capability detection are in aider/models.py. The Model class stores per-model settings (context window, edit format support, cost estimates) and is the canonical reference for Chapter 6's coverage of model selection. The get_model factory function shows how Aider resolves model names and applies default configurations.