README.md

---
title: ZeroGPU Batch Embedder
emoji: 🚀
colorFrom: blue
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
short_description: On-demand H200 batch embeddings for the RAG benchmark
---

# ZeroGPU batch embedder

Free H200 batch embedding for the GoodKnowledge RAG benchmark. Offloads the corpus-embedding
step (the local-Metal bottleneck) so the **full 512k EnterpriseRAG-Bench** is runnable.

## Deploy
1. Create a **ZeroGPU** Space (requires HF **PRO**), Gradio SDK.
2. Add **`HF_TOKEN`** as a Space secret (embeddinggemma is gated).
3. Push `app.py` + `requirements.txt`, set **Hardware → ZeroGPU** in Space settings.

## Use from the bench harness
Set the embedder to `hf-zerogpu:<your-username>/<space-name>` (see `_zerogpu_embed` in
`experiments/bench/bench_sweep.py`). The harness batches texts to fit the 60–120 s per-call
limit and caches the returned vectors. PRO = 25 min H200/day; 512k docs ≈ 2–4 min of GPU.