Behind the scenes · October 14, 2026 · 2 min read
How we find the right moment in hundreds of hours of audio
Keyword search misses meaning, and meaning-based search misses names. We use both, then re-rank.
A show with 300 episodes has around 300 hours of speech. Finding the two minutes that answer your question is the whole game. If we pick the wrong passages, even the best model will write a confident, wrong answer.
Two kinds of search
Meaning-based search turns your question and every passage into vectors, then finds passages that mean something similar. It knows that "trouble falling asleep" and "insomnia" are related. It's weaker on exact names, numbers and rare terms.
Keyword search is the classic kind. It's excellent at "what did they say about Huberman?" and bad at paraphrase.
We run both for every question and merge the two ranked lists with a simple, robust method called Reciprocal Rank Fusion. A passage that ranks well in either list makes the shortlist.
A second, closer read
The shortlist then goes to a re-ranker: a model that reads the question and each passage together and scores how well the passage answers it. We keep the best handful.
Small notes, big gains
Before indexing, each passage gets a one-line note describing what it's about in the context of its episode. A passage that just says "yeah, and that's why I stopped" becomes findable once its note says it's about quitting caffeine.
Measured, not guessed
Every change to this pipeline is tested against a set of real questions with known answers, including questions the show never covered. If a change doesn't improve the numbers, it doesn't ship.