Topic Segmentation: The Next Step

25 Jun, 2025

If you’ve been following along with my journey rebuilding https://talktothegame.felo5.com/ as a usable knowledge system — not just a transcript archive — here’s what’s next:

I’m adding topic segmentation via entity clustering.

Let me break it down, because this part might be one valuable shifts in how this system works.

What I’ve Built So Far

Right now, I’ve got a working RAG system. That means:

• I process each episode of Alex Hormozi’s podcast
• I chunk it into small, meaningful blocks
• I store those chunks in Supabase with embeddings
• I run speaker diarization, utterance segmentation, and even sentiment scoring
• And I enrich each chunk with the entities mentioned (people, companies, products, abstract ideas)

It’s already pretty solid. But something’s missing…

Why “Chunks” Aren’t Enough

Each chunk is useful in isolation.

But the truth is, Alex doesn’t speak in “chunks.”
He speaks in arcs. In themes.
One minute he’s talking about hiring, the next it’s pricing, and then he’s deep in some emotional story about grit.

And right now? My system doesn’t know where those transitions happen.

What I’m Building Next: Topic Segments

Instead of treating all chunks equally, I’m grouping them into semantic blocks:

Segments of an episode that focus on a single topic are automatically detected using the entities mentioned.

Here’s how it works:

1. Every chunk already has a list of entities
2. I look at the pattern of entities over time
3. When a new “cluster” of entities shows up consistently, I start a new topic segment
4. I give that segment a name (e.g. “Sales Hiring” or “Retention Strategy”)
5. And now I know:
   • Where the topic starts
   • What ideas dominate it
   • What to summarize together
   • What to skip or surface in a playlist

Why It Matters

• I won’t just “search” anymore, I’ll browse by concept
• I’ll be able to show “every time Alex talked about hiring” — grouped, not scattered
• It sets up everything else: insight scoring, quote detection, smart timelines, even agent routing

Why I’m Doing This Now

It’s the natural next step.
The RAG engine is humming. The data is enriched.
Entities are flowing in.
So instead of waiting to “do this later,” I’m wiring it into the enrichment pipeline right now — per episode.

Because this system isn't about transcripts.
It’s about turning raw speech into reusable thought.

What’s Next?

Once topic segments are live:

• I’ll start tracking how ideas evolve across episodes
• I’ll explore idea density and quote-worthiness inside each theme
• I’ll let users explore Alex’s mental model by domain: pricing, hiring, leverage, lead gen

All automated.
All tied to the authentic voice of the creator.
All queryable by meaning.

This is how I will outlearn every static archive on the internet.
This is how I’m turning a podcast into a living brain.

– Benoit Meunier