README.md

---
title: ReviewSearch
emoji: 🔍
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: mit
short_description: Semantic hybrid search over 206k ML/NLP peer reviews
---

# ReviewSearch

Semantic search over **205,988 ML/NLP peer reviews** (ICLR, NeurIPS, ICML, COLM, …). Hybrid retrieval:
a fine-tuned dense encoder (embeddinggemma-300m) fused with an inference-free learned-sparse encoder
(SPLADE) via reciprocal-rank fusion (k=10, 1:1), served on CPU over Qdrant.

- **Dense model:** [yjoonjang/reviewsearch-dense](https://huggingface.co/yjoonjang/reviewsearch-dense)
- **Sparse model:** [yjoonjang/reviewsearch-sparse](https://huggingface.co/yjoonjang/reviewsearch-sparse)
- **Index snapshot:** [yjoonjang/reviewsearch-index](https://huggingface.co/datasets/yjoonjang/reviewsearch-index)

## How it runs
On boot the container downloads the Qdrant snapshot from the dataset repo, recovers the collection,
and serves the API + React SPA. The dense/sparse models are pulled from the Hub on first use.

**Note:** free CPU Spaces sleep when idle; the first request after a cold start takes a few minutes
(snapshot download + model load), then queries run in ~tens of ms. ndcg@10 ≈ 0.258 on the judged gold
(hybrid, full-fidelity index).