21 August 2026
You don't search for the word you typed — you search for the word you remember, and those are rarely the same thing. Keyword search over meeting transcripts fails because it matches strings, not meaning; semantic search matches meaning, so it can find the moment someone said "forty cents a seat" even though you searched "pricing." What makes the result usable isn't the match itself — it's the date, the speaker, and enough surrounding speech to tell you whether it's actually the meeting you're thinking of.
A weekly vendor call runs for nine months and you've sat through thirty-nine of them. Somewhere in there, pricing came up. You remember the shape of the conversation — someone pushed back, a number got said out loud — but not which Tuesday it was, and not the words used to say it. That's the actual search problem: not "where is this," but "where is this thing I can describe but can't quote."
Keyword search tokenizes a transcript and matches literal strings. Type "pricing" and it looks for the token "pricing" — not "the number we talked about," not "what we'd charge them," not "the floor." People rarely use the word you'd use to describe the topic later; they use whatever phrase was natural in the moment, and that phrase is usually more specific and less searchable than the category you'd file it under.
The worse part is the silence. A keyword search that finds nothing looks identical whether the topic was never discussed or was discussed in different words. There's no signal telling you it's a language problem, not an absence.
Semantic search converts chunks of transcript into vectors that represent meaning, then converts your query the same way and ranks passages by how close the vectors sit — not by shared letters. Say "pricing" and it can surface a passage where nobody said that word, because "forty cents a seat," a currency figure, and a vendor name sitting next to a renewal date cluster near the same meaning. AVAY runs this over every past meeting a person has attended, so a query typed mid-call can pull a line said months earlier without anyone having tagged or filed it.
A good result is not a transcript dump. It's a passage — a handful of lines before and after the match, the speaker's name, the meeting title, and the date — enough to decide in three seconds whether this is the one, without opening the full recording.
Semantic search ranks by closeness in meaning, and closeness isn't always correctness. A passage about the wrong vendor's pricing can rank near the one you want if the topic and phrasing are similar enough. Negation is a known weak spot too — "we are not doing forty cents" and "we're doing forty cents" can land close together in meaning space even though they say opposite things.
It also only works on what got transcribed. A meeting where the recording was saved locally and never uploaded, or a side conversation that happened off-call, leaves nothing to search. The result list is only as complete as the meetings that actually ran through a transcript.
| Keyword search | Semantic search | |
|---|---|---|
| Matches on | the exact word or phrase you type | the meaning of what you describe |
| Best case | you remember the exact term someone used | you remember the topic, not the wording |
| Worst case | word was never said, returns nothing, looks empty | near-miss on the same topic surfaces the wrong meeting |
| Effort across 40 recordings | guess terms and scrub each transcript by hand | one query, ranked candidates across all of them |
That's the case semantic search is built for. It matches the meaning of your query against the meaning of what was said, so a search for "pricing" can surface a line where someone only said a dollar figure and a vendor's name, with no mention of the word pricing at all.
A search like this should cover every past meeting a person has attended, not just one recurring call. AVAY's search runs across all of it, so a query typed during today's call can pull up something said in an unrelated meeting six months ago.
No. Semantic search runs over transcript text, so a meeting with no transcript — because it wasn't recorded, or the recording sits only on someone's laptop and was never processed — has nothing for the search to reach.
The ranking gets you close; the surrounding context is what actually decides it. Check the date against what you remember, the speaker against who you think said it, and the sentence before and after the match — that's usually enough to know within a few seconds.
Not reliably. A sentence and its negation can land close together in meaning, since they're about the same topic in similar words, so it's worth reading the actual line rather than trusting the match blind for anything that hinges on a yes or no.
The word you remember is rarely the word that was said — search for the idea instead, and check the date and the sentence around the match before you trust it.
AVAY is a video meeting platform that transcribes the call itself — no bot joins, because there is nothing to join. Start one at avay.ai, read how each part works in the documentation, or see what it costs.