README.md

---
title: V-SPLADE Quality Document Retrieval
emoji: 🔍
colorFrom: pink
colorTo: indigo
sdk: gradio
sdk_version: 6.19.0
app_file: app.py
short_description: Visual document retrieval via sparse lexical vectors
python_version: "3.12"
startup_duration_timeout: 30m
---

# V-SPLADE Quality — Visual Document Retrieval

This Space demonstrates **V-SPLADE Quality** ([naver/v-splade-quality](https://huggingface.co/naver/v-splade-quality)),
a 0.25B inference-free sparse retriever for visual document retrieval.

Upload a document page image and enter text queries to see:
- The **top activated vocabulary tokens** — the sparse lexical representation of the image
- **Similarity scores** between each query and the document, with the top contributing tokens

V-SPLADE encodes document pages directly into sparse vocabulary vectors without OCR or captioning,
enabling retrieval 20× faster than caption-based pipelines.