modaic-ai

modaic-ai / gepa-viz

Public

Interactive live visualizer for gepa runs

35
1
89% credibility
Found May 24, 2026 at 142 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
TypeScript
AI Summary

gepa-viz is a live visualization tool for GEPA prompt optimization runs. It consists of two parts: a Python component that hooks into prompt optimization processes to record each candidate prompt and its evaluation results, and a web dashboard that renders this data as an interactive force-directed graph. Users can watch their prompts evolve in real-time, seeing which variations were accepted (shown as colorful donuts) versus rejected (shown as grey nodes), and click into any candidate to explore the prompt text, changes from the parent version, and detailed per-example feedback. The tool helps researchers and developers understand and debug their prompt optimization journeys.

How It Works

1
🔍 You learn about prompt optimization

You're working on improving AI prompts and hear about GEPA, a system that automatically evolves prompts to find the best ones.

2
📦 You install the visualization tool

You add gepa-viz to your project so you can watch your prompts evolve instead of staring at logs.

3
🔌 You connect the visualizer to your optimization

You add a simple line of code that tells GEPA to share its progress with the visualization tool.

4
🚀 You start your optimization run

Your AI starts testing different prompt variations, and everything is being recorded automatically.

5
🌐 You open the dashboard

A browser window opens showing a live graph that grows as your prompts evolve, node by node.

6
You explore the results your way
👆
Click any node

See the exact prompt text, how it changed from the parent, and detailed feedback for each test case.

🔲
Hover over connections

See why a proposal was rejected and what feedback the AI received that led to the next attempt.

✨ You find your best prompt

The visualization shows you the entire journey of prompts that were tried, what worked, what didn't, and the winning solution on the pareto frontier.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 142 to 35 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is gepa-viz?

gepa-viz is a live visualization tool for GEPA prompt-optimization runs. It renders the candidate tree as a force-directed graph, letting you watch prompts evolve over a pareto frontier in real time. The frontend is a React SPA with D3 for the graph rendering, while a Python package provides a callback you drop into your existing GEPA or DSPy scripts. You get a browser dashboard at localhost:5151 that updates node-by-node as the optimizer accepts and rejects proposals. Clicking any node reveals the prompt diff, per-example feedback from the reflection minibatch, and a clickable pixel grid showing the pareto frontier.

Why is it gaining traction?

The hook is watching your prompts evolve visually instead of parsing logs. The graph shows accepted candidates as donuts with green/red segments per example score, while rejected proposals appear as small grey nodes you can hover to see what feedback killed them. The pareto grid lets you drill into any example and compare prediction versus ground truth side-by-side. It's a single pip install and a CLI command away from having a real-time window into your optimization loop.

Who should use this?

Developers using GEPA or DSPy for prompt engineering who want visibility into why the optimizer accepts or rejects candidates. If you're iterating on prompts and finding the trial-and-error cycle opaque, this gives you the feedback loop in a visual form. Researchers benchmarking prompt strategies will appreciate the per-example breakdown. It's less useful if you just want final outputs without needing to understand the optimization process itself.

Verdict

gepa-viz fills a real gap for GEPA users who need to understand their optimization runs, but with 35 stars and a 0.8999999761581421% credibility score, it's early-stage and under-documented. The core functionality works, the CLI is simple, and the Python integration is clean. Just don't expect polished docs or a large community yet. Worth installing if you're actively using GEPA and want to see what's happening inside the black box.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.