Search without saying the word.

The clip you are looking for almost never contains the word you would type to find it. A speaker explains forgiveness for four minutes and never says the word. A founder describes the thing that nearly sank the company without once saying "crisis." The idea is in the room. The keyword is not.

Keyword search was built for documents — text that announces its own subject. Transcripts don't behave that way. People speak around an idea, circle it, arrive at it sideways. They reach for a metaphor instead of the term. They switch languages mid-sentence. The most important sixty seconds in an interview are often the ones where nobody states the headline out loud.

The failure is silent, which is what makes it expensive. A keyword search doesn't tell you it missed. It returns nothing, or the wrong nothing, and you conclude the moment isn't there. Hours of footage get written off as empty because the index only ever knew how to look for letters.

01Describing, not recalling

So we built search to match on meaning, not spelling. You describe what you remember — the gist, the situation, the feeling — and the system returns the moments that mean that, whether or not the words line up. Ask for "the part where she talks about starting over" and you get it, even though she said "I had to begin again from nothing" and never used your phrase.

This is a smaller change than it sounds, and a larger one. Smaller, because the interface is still a search box. Larger, because it removes the one requirement that made archives feel like work: you no longer have to remember the exact words to retrieve the exact moment. You only have to remember that the moment exists.

You no longer have to remember what was said. You only have to remember what it was about.

There is a quiet discipline behind this. Meaning-based retrieval can drift — it can hand back things that are vaguely related instead of actually relevant — and a vague answer is worse than no answer, because it costs you the time to check it. So every result points to a timestamp in a source file. You see the moment, not a summary of the moment. If the match is wrong, you know in a second. The system earns trust by being checkable, not by being confident.

It stays honest about language, too. An idea expressed in one tongue should surface when the question is asked in another, and the moment that comes back should be the real one — the speaker's own words at the second they said them — not a paraphrase that smoothed the edges off. Retrieval's job is to find. It is not licensed to rewrite.

What changes, once asking gets this cheap, is how people use their own footage. They stop rationing questions. They stop scrubbing a timeline hoping to stumble onto the right minute. They begin interrogating the archive the way you would a colleague who watched everything — where did this come up, show me the moment it turned, find the bit I am half-remembering. The archive answers in the language of the question, not the language of the metadata.