DSPyground

An open-source prompt optimization harness powered by GEPA. Install directly into your existing AI SDK agent repo, import your tools and prompts for 1:1 environment portability, and align agent behavior through iterative sampling and optimization—delivering an optimized prompt as your final artifact. Built for agentic loops.

Key Features

• Bootstrap with a Basic Prompt — Start with any simple prompt—no complex setup required. DSPyground will help you evolve it into a production-ready system prompt.
• Port Your Agent Environment — Use a simple config file to import your existing AI SDK prompts and tools—seamlessly recreate your agent environment for optimization.
• Multi-Dimensional Metrics — Optimize across 5 key dimensions: Tone (communication style), Accuracy (correctness), Efficiency (tool usage), Tool Accuracy (right tools), and Guardrails (safety compliance).

Quick Start

Prerequisites

• Node.js 18+
• AI Gateway API key (create one in the AI Gateway dashboard)

Installation

# Using npm
npm install -g dspyground

# Or using pnpm
pnpm add -g dspyground

Setup and Start

# Initialize DSPyground in your project
npx dspyground init

# Start the dev server
npx dspyground dev

The app will open at http://localhost:3000.

Note: DSPyground bundles all required dependencies. If you already have ai and zod in your project, it will use your versions to avoid conflicts. Otherwise, it uses its bundled versions.

Configuration

Edit dspyground.config.ts to configure your agent environment. All configuration is centralized in this file:

import { tool } from 'ai'
import { z } from 'zod'
// Import your existing tools
import { myCustomTool } from './src/lib/tools'

export default {
  // Your AI SDK tools
  tools: {
    myCustomTool,
    // or define new ones inline
  },

  // System prompt for your agent
  systemPrompt: `You are a helpful assistant...`,

  // Optional: Zod schema for structured output mode
  schema: z.object({
    response: z.string(),
    sentiment: z.enum(['positive', 'negative', 'neutral'])
  }),

  // Preferences - optimization and chat settings
  preferences: {
    selectedModel: 'openai/gpt-4o-mini',      // Model for interactive chat
    useStructuredOutput: false,               // Enable structured output in chat
    optimizationModel: 'openai/gpt-4o-mini',  // Model to optimize prompts for
    reflectionModel: 'openai/gpt-4o',         // Model for evaluation (judge)
    batchSize: 3,                             // Samples per iteration
    numRollouts: 10,                          // Number of optimization iterations
    selectedMetrics: ['accuracy'],            // Metrics to optimize for
    optimizeStructuredOutput: false           // Use structured output during optimization
  },

  // Metrics evaluation configuration
  metricsPrompt: {
    evaluation_instructions: 'You are an expert AI evaluator...',
    dimensions: {
      accuracy: {
        name: 'Accuracy',
        description: 'Is the information correct?',
        weight: 1.0
      },
      // Add more dimensions...
    }
  }
}

Configuration automatically reloads when you modify the file—no server restart needed!

Environment Setup

Create a .env file in your project root:

AI_GATEWAY_API_KEY=your_api_key_here

# Optional: For voice feedback feature (press & hold space bar in feedback dialog)
OPENAI_API_KEY=your_openai_api_key_here
OPENAI_BASE_URL=https://api.openai.com/v1  # Optional: Custom OpenAI-compatible endpoint

The AI_GATEWAY_API_KEY will be used by DSPyground to access AI models through AI Gateway. Follow the getting started guide to create your API key.

Voice Feedback (Optional):

• OPENAI_API_KEY: Required for voice feedback feature. Allows you to record voice feedback in the evaluation dialog by pressing and holding the space bar. Uses OpenAI's Whisper for transcription.
• OPENAI_BASE_URL: Optional. Set this if you want to use a custom OpenAI-compatible endpoint (e.g., Azure OpenAI). Defaults to https://api.openai.com/v1.

Note: All data is stored locally in .dspyground/data/ within your project. Add .dspyground/ to your .gitignore (automatically done during init).

How It Works

DSPyground follows a simple 3-step workflow:

1. Install and Port Your Agent

Install DSPyground in your repo and import your existing AI SDK tools and prompts for 1:1 environment portability. Use dspyground.config.ts to configure your agent environment.

2. Chat and Sample Trajectories

Interact with your agent and collect trajectory samples that demonstrate your desired behavior:

• Start with your system prompt defined in dspyground.config.ts
• Chat with the AI to create different scenarios and test your agent
• Save samples with feedback: Click the + button to save conversation turns as test samples
  • Give positive feedback for good responses (these become reference examples)
  • Give negative feedback for bad responses (these guide what to avoid)
• Organize with Sample Groups: Create groups like "Tone Tests", "Tool Usage", "Safety Tests"

3. Optimize

Run GEPA optimization to generate a refined prompt aligned with your sampled behaviors. Click "Optimize" to start the automated prompt improvement process.

The Modified GEPA Algorithm

Our implementation extends the traditional GEPA (Genetic-Pareto Evolutionary Algorithm) with several key modifications:

Core Improvements:

• Reflection-Based Scoring: Uses LLM-as-a-judge to evaluate trajectories across multiple dimensions
• Multi-Metric Optimization: Tracks 5 dimensions simultaneously (tone, accuracy, efficiency, tool_accuracy, guardrails)
• Dual Feedback Learning: Handles both positive examples (reference quality) and negative examples (patterns to avoid)
• Configurable Metrics: Customize evaluation dimensions via data/metrics-prompt.json
• Real-Time Streaming: Watch sample generation and evaluation as they happen

How It Works:

1. Initialization: Evaluates your seed prompt against a random batch of samples
2. Iteration Loop (for N rollouts):
   • Select best prompt from Pareto frontier
   • Sample random batch from your collected samples
   • Generate trajectories using current prompt
   • Evaluate each with reflection model (LLM-as-judge)
   • Synthesize feedback and improve prompt
   • Test improved prompt on same batch
   • Accept if better; update Pareto frontier
3. Pareto Frontier: Maintains set of non-dominated solutions across all metrics
4. Best Selection: Returns prompt with highest overall score

Key Differences from Standard GEPA:

• Evaluates on full conversational trajectories, not just final responses
• Uses structured output (Zod schemas) for consistent metric scoring
• Supports tool-calling agents with efficiency and tool accuracy metrics
• Streams progress for real-time monitoring

4. Results & History

• Run history stored in .dspyground/data/runs.json with:
  • All candidate prompts (accepted and rejected)
  • Scores and metrics for each iteration
  • Sample IDs used during optimization
  • Pareto frontier evolution
• View in History tab: See score progression and prompt evolution
• Copy optimized prompt from history and update your dspyground.config.ts

Configuration Reference

All configuration lives in dspyground.config.ts:

Core Settings

• tools: Your AI SDK tools (imported from your codebase or defined inline)
• systemPrompt: Base system prompt for your agent (defines agent behavior and personality)

Optional Settings

• schema: Zod schema for structured output mode (enables JSON extraction, classification, etc.)

Preferences

• selectedModel: Model used for interactive chat/testing in the UI
• optimizationModel: Model to generate responses during optimization (the model you're optimizing for)
• reflectionModel: Model for evaluation/judgment (typically more capable, acts as the "critic")
• useStructuredOutput: Enable structured output in chat interface
• optimizeStructuredOutput: Use structured output during optimization
• batchSize: Number of samples per optimization iteration (default: 3)
• numRollouts: Number of optimization iterations (default: 10)
• selectedMetrics: Array of metrics to optimize for (e.g., ['accuracy', 'tone'])

Metrics Configuration

• evaluation_instructions: Base instructions for the evaluation LLM
• dimensions: Define custom evaluation metrics with:
  • name: Display name for the metric
  • description: What this metric measures
  • weight: Importance weight (default: 1.0)
• positive_feedback_instruction: How to handle positive examples
• negative_feedback_instruction: How to handle negative examples
• comparison_positive: Comparison criteria for positive samples
• comparison_negative: Comparison criteria for negative samples

Voice Feedback Configuration (Optional)

• voiceFeedback.enabled: Enable/disable voice feedback feature (default: true)
• voiceFeedback.transcriptionModel: OpenAI Whisper model for transcription (default: 'whisper-1' — only Whisper supported)
• voiceFeedback.extractionModel: Model to extract rating and feedback from transcript (default: 'openai/gpt-4o-mini')

Note: Voice feedback requires OPENAI_API_KEY in your .env file. Press and hold space bar in the feedback dialog to record voice feedback.

Additional Features

• Structured Output Mode — Defin