How it works: two phases
All OpenAI does is turn text into numbers, or write text. The matching happens in our own database.
When you upload a meeting (happens once)
Transcript
→
Split into chunks
~200 words each (a few speaker turns).
→
OpenAI
Turn each chunk into numbers
1536 numbers = a “meaning fingerprint”.
→
Our database
Store chunks + numbers
Also, once: OpenAI reads the whole transcript → the meeting’s Decisions & Action items.
When you ask a question (every time)
Your question
→
OpenAI
Turn it into numbers
→
Our database
Find the 6 closest chunks
Nearest “meaning fingerprints”.
→
OpenAI
Write the answer
From those chunks, with citations.
Embedding = turning words into numbers that capture meaning. At question time our database — not OpenAI — compares the numbers to find the best matches.