How Market Bubble Search works

Every episode, indexed by meaning — and answered only from what was actually said.

1

Transcribe every broadcast

18 shows — every episode from #1 to #18, across 38 files: 25 YouTube uploads and 13 live X broadcasts, deduped down to the show. 66 hours, 68,525 lines.

Whisper, run locally
2

Cut into overlapping passages

3,107 windows. They overlap, so an answer that straddles a boundary is still findable.

every line keeps its timestamp
3

Index by meaning, not keywords

Each passage becomes a 1,024-dimension vector, so "why is eth done" finds it even if nobody said those words.

voyage-3.5 → Pinecone
4

Index the rare words exactly

19,449 names and numbers matched literally. "54 million" barely moves a vector, but it is the question.

the half meaning-search misses
5

Narrow to what actually answers it

Both routes feed 12 candidates into a reranker, which keeps the 6 that genuinely bear on the question.

rerank-2.5-lite
6

Answer from those six passages only

Claude writes the answer and cites the timestamp of the line it used — the second you can click.

claude-haiku-4-5

If it wasn't said on the show, it says so.

No outside knowledge. An answer with no citation is never posted — the timestamp is the part you can check.