Prompt Evolution Engine

Applies evolutionary algorithms to automatically optimize LLM prompts. Define a task, provide test cases (or let the system generate them), and watch evolution find a better prompt — generation by generation.

Based on EvoPrompt (ICLR 2024) and OPRO (Google DeepMind).

How It Works

1. Seed a population of prompts from your input + LLM-generated variations
2. Evaluate each prompt against test cases using an LLM-as-Judge (scored 0.0-1.0)
3. Select the fittest via tournament selection
4. Crossover + Mutate to create the next generation (7 mutation operators, section-aware crossover)
5. Repeat with adaptive mutation and elite re-evaluation
6. Result — the best prompt with fitness metrics and full genealogy DAG

Features

Quick Setup — structured 6-field prompt builder

Define your prompt using a guided framework: Persona, Task, Steps, Context & Constraints, Goal, and Output Format. The system auto-generates diverse test cases and assembles everything into an optimized prompt.

Advanced configuration

Full control over population size, generations, mutation rate, elite count, evaluation method, crossover strategy, and 7 mutation operators. Add custom seed prompts or let the system generate the initial population.

Multi-provider support

Run locally with Ollama (zero cost, no rate limits), or connect to Google AI Studio or OpenRouter for cloud inference. Select compute backend (Auto/GPU/CPU/Hybrid) and test connection before starting.

Real-time fitness tracking

Watch the population improve in real-time. The fitness chart tracks best, mean, and worst scores across all generations.

Prompt genealogy DAG

Interactive directed acyclic graph showing the full evolutionary lineage — which prompts were seeds, which came from crossover or mutation, and how elites carried forward. Click any node to inspect the prompt.

Generation-by-generation summary

Detailed table showing best, mean, and worst fitness for each generation with population counts.

Run history

Browse, compare, and revisit all past evolution runs with status, fitness scores, and timing.

Tech Stack

Next.js 16 | React 19 | TypeScript | Tailwind CSS 4 | shadcn/ui | SQLite + Drizzle ORM | Zustand | Recharts | React Flow | Zod | Vitest

Quick Start

Prerequisites

• Node.js 20+
• npm
• Ollama (for local inference) or a Google AI Studio / OpenRouter API key

Setup

git clone https://github.com/Blizzeq/Prompt-Evolution-Engine.git
cd Prompt-Evolution-Engine

npm install
cp .env.local.example .env.local   # configure your provider
npx drizzle-kit migrate
npm run dev

Open http://localhost:3000.

The SQLite database is created automatically at ./data/evolution.sqlite by default. Override it with DB_PATH in .env.local if needed.

Using Ollama (local, recommended)

ollama serve
ollama pull gemma4    # Gemma 4 26B — MoE, activates 4B params/token

Using cloud providers

Set GOOGLE_AI_API_KEY or OPENROUTER_API_KEY in .env.local.

Research References

• EvoPrompt — LLMs with Evolutionary Algorithms (ICLR 2024)
• OPRO — Optimization by PROmpting (Google DeepMind)
• DSPy MIPROv2 — Bayesian prompt optimization (Stanford NLP)
• GAAPO — GA-based Automated Prompt Optimization

License

MIT