GEPA Integration
Graph Enhanced Prompting Algorithm - Revolutionary reflective prompt optimization that leverages Large Language Models' ability to analyze and critique their own behavior, achieving dramatic performance improvements with minimal training data through domain-specific textual feedback.
How SuperOptiX Optimizes Any Framework with GEPA
SuperOptiX uses a single, universal GEPA workflow to optimize agents across frameworks like DSPy, OpenAI SDK, CrewAI, Pydantic AI, Google ADK, Microsoft, and DeepAgents. You keep one SuperSpec, and simply pick the target framework in configuration.
Define your agent once in SuperSpec, compile to any supported framework and let GEPA optimize.
Switch between DSPy, Pydantic AI, CrewAI, and more just by changing a config field.
Metrics, budgets, and reflection models stay the same across frameworks.
Consistent GEPA traces and metrics no matter which runtime you use.
Universal GEPA Configuration (Framework-Agnostic)
Same GEPA block, different framework and runtime targets
One Workflow, Different Framework Targets
Same GEPA commands, just switch --framework
Default `legacy` + Experimental `optimize_anything`
SuperOptiX defaults to the stable legacy GEPA API mode and supports explicit opt-in to the experimental optimize_anything mode.
Mode Behavior
Docs-aligned behavior used by SuperOptiX CLI and playbooks
legacy is the default mode and preserves backward-compatible behavior.optimize_anything is available as an explicit, experimental opt-in.optimize_everything is accepted as an alias and mapped internally to optimize_anything.optimize_anything is unavailable in the installed GEPA version, SuperOptiX falls back to legacy automatically.CLI Usage
Set mode explicitly with --gepa-api
Playbook Configuration
Enable experimental mode with gepa_api in optimizer params
spec:
optimization:
optimizer:
name: GEPA
params:
auto: light
reflection_lm: llama3.1:8b
gepa_api: optimize_anything # optional; default is legacyResearch Foundation
GEPA is based on cutting-edge research in prompt optimization and reflective learning, demonstrating that reflective prompt evolution can outperform reinforcement learning approaches while requiring significantly less computational resources and training data.
The key innovation is GEPA's ability to utilize domain-specific textual feedback (compiler errors, medical guidelines, security advisories) rather than just numeric scores, enabling more targeted and effective optimization.
GEPA Architecture
Reflective Analysis
A reflection LM analyzes agent trajectories to identify specific failure modes and improvement opportunities
Prompt Evolution
New prompt candidates are generated based on reflective insights and domain-specific feedback
Graph Construction
A tree of evolved prompts is built, with Pareto-aware selection preserving improvements
Iterative Refinement
The process repeats, accumulating improvements over multiple generations
Why GEPA is Revolutionary
GEPA brings breakthrough capabilities that set it apart from traditional optimization approaches
Sample Efficiency
GEPA often achieves significant improvements with far fewer training examples than traditional methods. Where other optimizers might need hundreds of examples, GEPA can improve performance with just 3-10 well-chosen scenarios.
Domain Adaptability
GEPA excels at incorporating domain-specific knowledge through textual feedback, making it particularly effective for specialized applications like mathematics, medicine, law, and security.
Interpretable Improvements
Unlike black-box optimization, GEPA generates human-readable prompt improvements that you can understand and validate.
Multi-Objective Optimization
GEPA can simultaneously optimize for multiple criteria (accuracy, safety, compliance) through its feedback system.
🚀 Interactive GEPA Demo
For the most up-to-date GEPA information and examples, use the official optimization guide and GEPA repository instead of older demo repos.
Complete Configuration Guide
Comprehensive GEPA configuration options with domain-specific setups and best practices
Basic GEPA Configuration
Standard GEPA setup with light budget for development
Quick Start Commands
Get started with GEPA in minutes
Available GEPA Agents
Pre-configured GEPA agents across multiple domains, each optimized for specific use cases with domain-specific feedback metrics
🧮 Mathematics & Analytics
Advanced Math GEPA Solver
Agent ID: advanced_math_gepa
Domain & Specialization
Advanced mathematical problem solving
Specializes in: Quadratic equations, calculus, geometry, algebraic reasoning
Key Features
Data Science GEPA
Agent ID: data_science_gepa
Domain & Specialization
Statistical analysis and machine learning
Specializes in: Data analysis, statistical inference, ML insights, hypothesis testing
Key Features
🏥 Healthcare & Medical
Medical Assistant GEPA
Agent ID: medical_assistant_gepa
Domain & Specialization
Clinical decision support and medical information
Specializes in: Medical knowledge synthesis, patient education, clinical reasoning
Key Features
⚖️ Legal & Compliance
Contract Analyzer GEPA
Agent ID: contract_analyzer_gepa
Domain & Specialization
Legal contract analysis and risk assessment
Specializes in: Contract review, risk identification, compliance verification
Key Features
💼 Enterprise & Finance
Enterprise Extractor GEPA
Agent ID: enterprise_extractor_gepa
Domain & Specialization
Enterprise document processing and information extraction
Specializes in: Multi-component analysis, structured data extraction, business intelligence
Key Features
🔒 Security & Privacy
Security Analyzer GEPA
Agent ID: security_analyzer_gepa
Domain & Specialization
Security vulnerability detection and code analysis
Specializes in: Vulnerability detection, secure coding practices, security assessment
Key Features
Privacy Delegate GEPA
Agent ID: privacy_delegate_gepa
Domain & Specialization
Privacy-preserving task delegation and data handling
Specializes in: Data anonymization, privacy compliance, secure information handling
Key Features
Performance Benchmarks
Real-world performance improvements across different domains with GEPA vs traditional optimization
| Domain | Baseline | GEPA Result | SIMBA Result | GEPA Advantage | Description |
|---|---|---|---|---|---|
| Mathematics | 60% | 95% | 90% | +35% | Quadratic equations and geometric problems with step-by-step reasoning |
| Document Analysis | 45% | 85% | 70% | +40% | Multi-component enterprise document processing and extraction |
| Security Analysis | 70% | 95% | 80% | +25% | Vulnerability detection with detailed remediation guidance |
| General Q&A | 80% | 88% | 85% | +8% | General-purpose question answering and reasoning tasks |
* Benchmarks based on SuperOptiX testing with llama3.1:8b on M4 Max (128GB). Results may vary based on hardware and configuration.
⚠️ Important: GEPA and Tool-Calling Agents
GEPA is not compatible with ReAct agents that use tool calling (Genies tier and above).This includes any agent with tool integration.
Why GEPA doesn't work with tool-calling agents:
Agent Tier Compatibility:
Complete GEPA Documentation
Explore comprehensive guides, API references, and practical examples
Ready to Transform Your Agents with GEPA?
Experience the power of reflective prompt optimization and see dramatic performance improvements with minimal training data. GEPA represents the future of agent optimization.