Multi-Model Workflows

Overview

Multi-model workflows use different AI models for different tasks, combining their strengths. A fast model might handle initial classification while a more capable model handles complex reasoning.

Why Use Multiple Models

Different models excel at different tasks. Combining them lets you optimize for multiple objectives simultaneously.

• Cost Optimization: Use expensive models only when needed. Route simple requests to cheaper models, saving 80%+ on routine tasks.
• Speed & Quality Balance: Fast models for latency-sensitive steps, powerful models for quality-critical steps. Get the best of both.
• Specialized Strengths: Some models excel at code, others at reasoning, others at creativity. Use the right tool for each job.
• Reliability: If one provider has issues, fall back to another. Reduces single points of failure.

Common Patterns

These are the most effective multi-model patterns used in production workflows.

Model Routing

Route requests to different models based on characteristics like complexity, topic, or user tier. ORCFLO provides two main routing mechanisms: If/Else nodes for rule-based routing and Criteria Check for intelligent AI-powered routing.

Input → Criteria Check / If-Else Node
        │
        ├── simple    → Claude Haiku 4.5   → Output
        ├── moderate  → Claude Sonnet 4.5  → Output
        └── complex   → Claude Opus 4.5    → Output

Cost savings: 60-80% compared to using Opus for everything

Routing Criteria

• Input length (short to fast model, long to capable model)
• Task type (classification to small, generation to large)
• Quality requirements (internal to small, customer-facing to large)
• User tier (free to economical, premium to best quality)
• Detected complexity (AI-assessed difficulty score)

Model Chaining

Chain multiple models in sequence where each builds on the previous model's output.

Step 1: Draft Generation (GPT-4o Mini - fast & cheap)
─────────────────────────────────────────────────
Task: "Write a first draft blog post about the provided topic"
Output: Rough draft with key points

Step 2: Quality Refinement (Claude Sonnet 4.5 - powerful)
─────────────────────────────────────────────────
Task: "Improve this draft. Fix any errors, improve
       flow, and make it more engaging."
Input: Output from draft node (automatically passed)
Output: Polished final version

Chain Examples

• Extract then Analyze: Fast model extracts structured data, powerful model performs complex analysis on the clean data.
• Translate then Localize: One model translates, another adapts cultural references and idiomatic expressions.
• Generate then Validate: One model generates content, another checks for accuracy, safety, or policy compliance.

Fallback Strategies

Use fallbacks to handle model failures or quality issues gracefully.

1. Try Haiku 4.5 (fast, cheap)
   └── If confidence < 0.8 or output seems poor
       └── 2. Retry with Sonnet 4.5 (powerful)
           └── If still failing
               └── 3. Use Opus 4.5 (most capable)

Most requests resolve at step 1, saving costs.
Complex cases automatically escalate.

Best Practices

• Keep routing logic simple - complex routing can negate cost savings
• Use the cheapest effective model for each step
• Test each model independently before combining
• Monitor per-model costs and quality metrics
• Have fallbacks for reliability, not just cost optimization
• Document which model is used for what and why
• Re-evaluate model choices as new models are released