# VideoDB

> An applied research lab for visual intelligence. Build intelligence for the visual world.

Data infrastructure for video, built to train and deploy visual AI.

Canonical: https://videodb.io · Latest: [VideoDB leads video retrieval benchmarks across 9,800 queries](https://arxiv.org/abs/2608.08075)

Backed and supported by Mango Capital, Strange Ventures, Untapped Capital, OVO Fund, UC Berkeley and IIIT Bangalore.

## 01 / What we do

We turn video into data you can use. We investigate where your model fails, find the data that fixes it, and bring scalable infrastructure and the expertise to go with it, so your team builds intelligence from video faster.

1. **Probe the test runs.** Test runs are annotated and indexed as they finish. Search agents go through each run, pull out the moments where the model failed and group them by cause, so the team sees which failure to work on first. Lab post: https://videodb.io/blog/robot-runs.md
2. **Find the episodes you already have.** Ask across your in-house datasets in plain language: same task, objects and lighting. Results come back as ranked episodes, with frames. VideoDB leads search benchmark across 9,800 queries: https://arxiv.org/abs/2608.08075
3. **Connect more sources when the archive runs out.** When the archive does not cover a failure, we bring in more: partner datasets, targeted collection, or live runs and cameras streamed into the same pipeline.
4. **Annotate, with human QC.** Any model runs as an analyzer over the episodes. Low-confidence labels go to human reviewers with the clip attached. 100,000 hours annotated in a few days for one customer. VideoDB leads on WGO-Bench: https://videodb.io/blog/wgo-bench-action-annotations.md
5. **Retrain and test again.** The new samples go into the next training run. The updated model runs the same tests on the same infrastructure, and the next round of runs is searchable as soon as it finishes.

All products: https://videodb.io/products.md

## How VideoDB works

VideoDB built a distributed data format for machine understanding of the visual world. It sits as a logical layer on top of your existing files and spreads the work across as many machines as you give it. Every video primitive in your workflow, from transcoding to search, runs on top. Three distributed processing stages write into the format (transcoding, frames, annotation) and three scalable serving stages read from it (retrieval, streaming, runtime). Use one API. Run in our cloud or yours. The full pipeline: https://videodb.io/platform.md

Self-serve: get an API key at https://console.videodb.io/auth, upload a file, index it, search it, play the evidence. Docs and quickstart: https://docs.videodb.io

## 02 / Research

Benchmarks, papers and engineering notes from the lab. Most of what we publish is measured on public benchmarks, with the numbers included.

- [Search over the Visual World](https://arxiv.org/abs/2608.08075): technical report. Recall@1 of 73.1% against 65.8% for a commercial video-native engine, over 9,834 queries from four public datasets (Aug 8, 2026).
- [What video retrieval benchmarks taught us about ground truth](https://videodb.io/blog/video-benchmark-ground-truth-ambiguity.md): a manual review of MSR-VTT, MSVD, VATEX, DiDeMo and QVHighlights, where the labelled window is narrower than the clips that match the query (Jul 24, 2026).
- [Can a VLM annotate a robot run?](https://videodb.io/blog/wgo-bench-action-annotations.md): all 100 WGO-Bench videos re-annotated. A VLM pipeline reaches 28.9% semantic F1 at IoU 0.5, against a 19.9% public baseline (Sep 8, 2026).

Papers and notes: https://videodb.io/blog.md

## 03 / Talk to us

The next wave of AI will understand the visual world. 10k+ developers on the platform. 250+ TB processed. [Talk to us](https://videodb.io/contact.md) · [Read the lab post](https://videodb.io/blog/robot-runs.md)

To see is to know.