Provider Integration¶

DeepFabric integrates with language model providers, enabling access to virtually any commercial or open-source language model through a unified interface. This abstraction layer provides consistent behavior across different providers while preserving access to provider-specific capabilities and optimizations.

The integration approach supports both simple single-provider workflows and sophisticated multi-provider strategies that optimize different components for cost, performance, or capability requirements.

Supported Providers¶

DeepFabric supports all model providers through consistent configuration patterns:

OpenAI provides GPT models including GPT-4, GPT-3.5, and specialized variants optimized for different use cases.

Anthropic offers Claude models with strong reasoning capabilities and detailed response generation.

Google includes Gemini models with multimodal capabilities and competitive performance characteristics.

Ollama enables local model execution with privacy benefits and no per-token costs.

Azure OpenAI provides enterprise-grade OpenAI model access with additional security and compliance features.

AWS Bedrock offers access to multiple model providers through a single enterprise-focused platform.

Authentication Patterns¶

Each provider uses environment variables following the {PROVIDER}_API_KEY pattern:

# OpenAI authentication
export OPENAI_API_KEY="your-openai-key"

# Anthropic authentication
export ANTHROPIC_API_KEY="your-anthropic-key"

# Google Gemini authentication
export GEMINI_API_KEY="your-gemini-key"

# Hugging Face (for dataset upload)
export HF_TOKEN="your-hf-token"

This consistent pattern simplifies multi-provider workflows and enables easy provider switching through environment configuration.

Configuration Syntax¶

Provider configuration uses the provider/model format throughout DeepFabric:

# OpenAI GPT-4
provider: "openai"
model: "gpt-4"

# Anthropic Claude
provider: "anthropic" 
model: "claude-3-opus"

# Local Ollama
provider: "ollama"
model: "mistral:latest"

# Google Gemini
provider: "gemini"
model: "gemini-pro"

This separation enables clear specification of both the provider infrastructure and the specific model within that provider's catalog.

Multi-Provider Workflows¶

DeepFabric supports using different providers for different components, enabling optimization strategies:

# Fast, cost-effective topic generation
topic_tree:
  provider: "openai"
  model: "gpt-3.5-turbo"

# High-quality dataset generation
data_engine:
  provider: "anthropic"
  model: "claude-3-opus"

# Different model for final dataset creation
dataset:
  creation:
    provider: "ollama"
    model: "mistral:latest"

This approach balances cost, speed, and quality by using appropriate models for each stage's specific requirements.

Provider-Specific Optimizations¶

Different providers have distinct characteristics that influence optimal usage patterns:

OpenAI Models excel at following instructions precisely and provide good balance of speed and quality. GPT-4 offers superior reasoning while GPT-3.5 provides faster, more economical processing.

Anthropic Claude demonstrates strong performance on complex reasoning tasks and produces detailed, well-structured responses ideal for educational content generation.

Ollama Local Models eliminate per-token costs and provide complete data privacy, making them suitable for sensitive content or high-volume generation with budget constraints.

Google Gemini offers competitive performance with potentially lower costs and good instruction-following capabilities.

Rate Limiting and Error Handling¶

DeepFabric includes comprehensive error handling for provider-specific issues:

Automatic Retry Logic handles temporary failures with exponential backoff, adapting to different providers' reliability characteristics.

Rate Limit Management respects provider-specific rate limits through configurable delays and batch size management.

Fallback Strategies can be configured to switch providers when primary providers experience issues or reach limits.

Cost Monitoring tracks usage patterns to help optimize provider selection for budget-conscious deployments.

Local Model Setup¶

Ollama provides completely local model execution with simple setup:

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

# Pull and run a model
ollama pull mistral:latest
ollama pull codellama:latest

# Verify model availability
ollama list

Local models require no API keys and provide unlimited usage, though they require sufficient local computational resources.

Provider Selection Strategy¶

Choose providers based on your specific requirements:

For Development and Testing use faster, less expensive models like GPT-3.5 or local Ollama models to iterate quickly on configuration and approach.

For Production Datasets consider higher-quality models like GPT-4 or Claude-3-Opus when output quality is paramount.

For Privacy-Sensitive Content use local Ollama models to maintain complete control over data processing.

For Cost Optimization mix providers strategically, using economical models for topic generation and premium models only for final dataset creation.

Performance Monitoring¶

Monitor provider performance through built-in metrics:

Response Times vary significantly between providers and models, affecting overall generation time.

Success Rates differ based on provider reliability and model capability for your specific use cases.

Cost Tracking helps optimize provider selection for budget-conscious projects.

Quality Assessment evaluates output quality across different providers to inform selection decisions.

Provider Migration Strategy

Start with a single provider for simplicity, then experiment with multi-provider approaches as you understand your specific requirements. Keep provider configurations in separate YAML files to enable easy switching for different use cases or cost optimization needs.