---
title: ElideDB Query by Example
emoji: 🎬
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# ElideDB — query by example

1,122 robot episodes, and not one word of text. Pick a few clips that show
the same kind of moment; the archive returns its own kind.

**Why this exists.** A text encoder scores *"open the drawer"* against
*"close the drawer"* at cosine 0.977, and a pooled video embedding scores a
clip against its own reversal at 1.000000. Description destroys the
distinction before search begins. A clip needs no description: it is an
instance of what it shows.

**What runs when you click.** Nothing loads. Every vector was computed when
the video was written, so a query is arithmetic over parquet columns —
about 250 ms over the whole archive. Six channels vote, weighted by how
tightly each one pulls *your* picks together relative to how it holds the
archive: a statistic taken from the query itself, with no labels and no
training. The panel shows those weights, because which model recognised
your pick is the interesting part of the answer.

**How many come back.** The count is a ceiling, not a target. Each pick is
held out in turn to see how deep its own kind ranks, and the set stops
where that evidence stops — returned means believed.

Thumbnails and clips are decoded from the stored video's own byte ranges on
demand. There is no preview cache anywhere in the store.

The text-query demo of the same engine is at
[SudharshanR/elidedb-demo](https://huggingface.co/spaces/SudharshanR/elidedb-demo).