ELIDEDB

Show your footage a moment. Get back every time it happened.

ElideDB is a memory for video. Ingest your recordings once. Then hand it a clip, a spill, a dropped part, a near miss, and it returns every other time something like that happened, ranked, with the span inside each recording, in about half a second. No labels. Nothing trained on your data. Works on footage it has never seen.

Try the live demo loads in seconds: nothing it ranks needs a model at query time

runs on your hardware; your video never leaves built on frozen open models from Meta and Google every number on this page is a measurement

elidedb

$ elidedb add    kitchen  /videos/robot_runs   # ingest, once
$ elidedb query  kitchen  /clips/spill.mp4     # ask with a clip

 #   match             span  source
 1   41.2%   12.5-  24.3s  run_0114/frames.mp4
 2   38.7%    0.0-  11.8s  run_0088/frames.mp4
 3   31.1%    3.2-  15.0s  run_0203/frames.mp4

3 hits in 554 ms

the problem

Your fleet records everything. Almost none of it is ever looked at again.

A gripper drops a part overnight. A test drive clips a cone. A customer reports the arm "did something weird" on Tuesday. You have the footage. Somewhere. In hours of recordings across a dozen cameras.

The property you care about, what happened and in what order, is exactly what text search and image search are worst at. To a model trained on captioned photographs, opening a drawer and closing it are nearly the same picture. The difference is in the order the frames arrive.

4 TB a day from one self-driving car; a test vehicle makes 11 to 152 TB. Ten vehicles on a shift produce 80 hours of footage a day. Watching it once is ten people doing nothing else.

$2,500 a question an engineer hour costs about $64 at U.S. median pay, and reviewing footage takes at least as long as the footage. One question over one week of recordings, before any real work begins.

800 hours to label one annotating a single hour of driving data takes about 800 person-hours, and the labels only answer the questions the schema anticipated. Every new question is a new labeling pass.

90% never analyzed IBM's estimate for data generated by devices, connected vehicles included. The footage sits there and the fleet keeps making the same mistake.

Sources for every figure are linked in the project README.

the fix

Ask your footage with footage. Get moments, not homework.

1Point it at your recordings Video goes in as it is. Ingest is one pass, about 14 minutes of one GPU per hour of video, resumable. Your files are never copied or modified.

2Show it a moment A clip from anywhere, or a slice of a longer video. A robot asks with its own live trace, which is already in the store, so the question costs nothing to ask.

3Get every time it happened Ranked recordings, each with the span where the match lies, in about half a second. Top result right 95% of the time on tasks the system had never seen.

you are probably thinking

The questions every engineering lead asks us

"We do not have budget for a labeling team."

It never asks for labels. Two frozen, off-the-shelf encoders and a time-alignment step. There are no ElideDB weights. Nothing is tuned to a dataset, so the same system works on kitchens, warehouses, and roads on the day it is installed.

"Models work on the demo data and fall apart on ours."

Ours was measured on a corpus where 91% of the tasks were never seen during setup, and it scores slightly higher on the unseen tasks than the familiar ones. There is no home corpus to be biased toward, because nothing was trained on one.

"Our footage cannot leave the building."

It does not have to. ElideDB runs on your hardware or in your cloud account, with no per-query API call to anyone. The store holds compact derived traces and a manifest pointing at your files. The video never moves.

"We run several sites. Will their data mix?"

No. One store per robot, site, or customer, and every statistic the search uses is computed inside the store. Two deployments cannot influence each other's results, by construction.

what that buys you

Half a second where you spend afternoons

95%

top result correct, on a held-out corpus where 91% of tasks were never seen during setup

554 ms

median query over a 3,556-recording store; a robot querying its own memory encodes nothing

labels, tags, or annotations you will ever write; zero parameters trained on your data

14 min

of one GPU to ingest an hour of video, once; about 0.8 GB of traces per video-hour

Where it is weak, stated plainly. Precision falls off past roughly the top 20 results. It is strong at "show me the closest matches" and not yet at "show me every instance": asked for all 35 of 35, it currently finds about 40%. That gap is the top of the roadmap, and we publish it rather than hide it.

faq

Fair questions

What footage does it work on?

Any video: robot fleets, dashcams, factory floors, lab rigs, drones. It was measured across simulated kitchens, real robot manipulation, road driving, and aerial footage with the same untouched system. Clips shorter than four seconds cannot be encoded.

Do we need GPUs?

For ingest, you want one: about 14 minutes of GPU per hour of video on Apple Silicon; CUDA is supported. A robot querying its own memory needs no encoder at all. Querying with an outside clip runs the encoders once per clip, on GPU or CPU.

Is there a language model making things up?

No. There is no generative model anywhere in the path. Two frozen vision encoders turn video into traces, and an elastic time alignment ranks them. Every result is evidence from your actual frames.

Can I search with words?

Not yet, by design. An earlier version searched by sentence and it taught us the problem: words erase what happened and in what order. That version is deprecated. A vocabulary layer that sits on top of the motion representation, rather than instead of it, is being optimised for release.

What happens to our raw files?

They are read once and never copied, moved, or modified. The store keeps compact derived traces, about 0.8 GB per hour of video, and a manifest pointing at your files. Deleting a store deletes only the traces.

Can we evaluate it on our own data before anything else?

That is the intended first step. The source is available for evaluation, testing, and research, with a manual that walks through ingesting your footage and scoring results against your own ground truth. Commercial use is licensed separately.

What does it cost?

We are onboarding early teams now and shaping pricing with them. The cost model underneath is simple: ingest is a one-time compute cost per hour of video, and every query after that is half a second on a machine you already own.

read the whole thing

Everything is written down

The projectwhat it is, the numbers, the weak spots Using itinstall, ingest your footage, query, evaluate on your data How it worksmoment retrieval step by step, and why nothing is trained Roadmaplimitations, the growth trajectory, how customer data compounds The storywhat was tried, what failed, why the design is what it is Experimentsevery approach tried and what it scored, including the dead ends

from the founder

I spent years at NYU building synchronized capture rigs: cameras, LiDAR, and microphones recording city streets for research. We got very good at collecting data. Then I watched every team, ours included, hit the same wall: the recordings pile up, and the only search tool anyone has is their own eyes.

ElideDB is the memory I wished those rigs had. We spent weeks trying to teach a model what "similar" means, and the thing that finally worked on footage it had never seen was refusing to teach it anything. It answers from your frames, in half a second, and it tells you where it is weak.

Sudharshan Ramesh founder, ElideDB

see it yourself

Ask real robot footage with real robot footage

The demo holds 1,122 robot manipulation recordings. Pick a few clips of a kind and it returns the rest of that kind. Nothing it ranks needs a model at query time, so it loads in seconds. Production deployments run on your own hardware, next to your data.

Open the live demo Get early access