Visualize a GEPA or DSPy prompt optimization run as an interactive node graph that updates live in your browser.
Click on rejected prompt nodes to read the feedback that caused each candidate to be discarded.
Compare prompt performance across test examples using the color-coded donut rings displayed on each graph node.
Modify the visualization front end by running the live-reload dev server without touching the Python package.
Requires an existing GEPA or DSPy optimization script to generate the run data that feeds the visualization.
gepa-viz is a live visualization tool for watching AI prompt optimization runs in your browser. It connects to GEPA, a prompt optimization library, and draws an interactive graph that updates in real time as the optimizer explores and tests different prompt candidates. The idea is to give you a visual window into a process that would otherwise be an opaque stream of terminal output. The graph works like a family tree of prompts. Each prompt the optimizer accepts appears as a donut-shaped node, and the colored ring around it shows how well that prompt performed on individual test examples (green for correct, red for wrong). Rejected prompt attempts show up as small grey nodes. Hovering over a grey node reveals the feedback that caused it to be discarded. Clicking any node opens a detail panel with the full prompt text, a diff showing what changed from the parent prompt, and a breakdown of how the optimizer evaluated it. Installation is a single pip command. Once installed, you add one callback object to your existing GEPA or DSPy optimization script, then run a separate terminal command to start a local web server. A browser tab opens automatically and the graph grows node by node as the run progresses. There is no Node.js or other runtime required to serve it. Under the hood the project is split into a Python package that captures the run data and writes it to a JSON file, and a browser front end built with React, D3, and Tailwind that reads and renders that file. The pre-built browser app is bundled inside the Python wheel, so end users never need to touch the front-end code. Developers who want to modify the visualization can run a live-reload dev server with a few provided commands.
← modaic-ai on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.