# Products · VideoDB

> Products built for teams processing PBs of video.

Each one runs as a managed package on our cloud or yours. Modular building blocks, so you can take the ones you need.

Canonical: https://videodb.io/products

## 01 / Video primitives: transcoding, frames and streaming, built for random access

Transcoding, frame tiling, real-time ingestion and high-throughput pipelines, built in house. Ask for one frame or one window and get exactly that, without re-encoding.

- **In-house transcoding**: any input format is transcoded once into a frame-addressable form. Seeking to a frame costs the same anywhere in the file.
- **Frame tiling**: frames from a time window are tiled into one image at the resolution you set, so a model reads the whole window in one call.
- **Frame-addressable streaming engine**: request any frame range of any episode and get it streamed, to a model or a player, in one call.
- **Real-time ingestion**: live cameras and screen recordings go through the same pipeline as uploaded files.
- **High-throughput pipelines**: 100,000 hours turned into scene-level samples in a few days, files left in place.

The architecture: https://videodb.io/platform.md · Docs: https://docs.videodb.io

## 02 / Machine annotation: segmentation and dense annotations of every hour you have, at scale

Run any model over your footage: ours, open-weight, or one you trained. Outputs go into a versioned index per model, so a better model can re-annotate everything later. Human QC through our partners for the cases the model is unsure about. 100,000+ hours in a few days.

- **Any model as an analyzer**: ours, open-weight, proprietary, or the one you trained last week. One versioned index per model; re-running a better model rewrites the index without touching the files.
- **Segmentation and dense annotation**: objects, hands, actions, on-screen text and speech, labelled at frame, scene or episode level.
- **Data sources and data providers**: S3 or GCS buckets, fleet uploads and footage from collection partners, all into one index.
- **Human QC, with partners**: low-confidence labels and rare events are sent to reviewers with the clip attached. Their answer is written back to the same index.

Annotation benchmark: https://videodb.io/blog/wgo-bench-action-annotations.md · Talk to us: https://videodb.io/contact.md

## 03 / Episode retrieval: deep search every episode by what happens in it

Build agents to find every episode you have, apply filters and aggregations. Get results as flexible-length clips and frames within 500 ms.

- **Measured on public datasets**: Recall@1 of 73.1% against 65.8% for a commercial video-native engine, over 9,834 queries from four public datasets. Paper: https://arxiv.org/abs/2608.08075
- **About 500 ms to the moment**: queries in plain English. Results come back as clips sized to the query, whether that is a two-second regrasp or a full shift.
- **From hits to training set**: filter and balance the results, then export a manifest your training code can load.
- **Agents with deep search**: deep-search, aggregation and watch agents, usable from any MCP client or the Python and Node SDKs.

Lab post: https://videodb.io/blog/robot-runs.md · Talk to us: https://videodb.io/contact.md

## 04 / Realtime ingestion: live cameras and streams, analyzed as they run

Alerts and reactive systems on live video. Live runs and cameras go through the same pipeline as the archive.

- **Search while it streams**: cameras, live streams and screen recordings are indexed while they run.
- **Alerts with the clip attached**: describe the event in plain language. Each alert carries the clip that triggered it.
- **Live robot runs**: a run is searchable while it is still happening, and stays in the archive when it ends.
- **Computer-use agents**: record agent sessions continuously and search them afterwards: what did the agent click before the error?

RTSP guide: https://videodb.io/blog/rtsp-ai-analysis.md · Book a demo: https://videodb.io/contact.md

## 05 / Connected to your robot data: reads MCAP, LeRobot and RLDS, writes manifests and PyTorch loaders

Your recordings stay where they are and stay the source of truth. VideoDB indexes the video streams inside them and links every result back to the episode, timestamp and camera.

- **Inputs**: MCAP, LeRobot, RLDS, RTSP, and files in S3, GCS or Azure. Simulator and world-model output too.
- **Outputs**: Parquet manifests, a PyTorch loader adapter, MCAP shards, evidence streams, scheduled agent results.
- **Provenance**: each result carries episode id, timestamp, camera, source type and model version.
- **Real and simulated**: real, simulated and generated video are tagged separately in the index and in every export.

Talk to us: https://videodb.io/contact.md