Building
An early Reso user once told me: "I have 200 voice notes and I can't remember what half of them are about."
I related immediately, because my own archive felt the same. Capture-focused tools make recording easy, then leave you drowning in data.
I wanted Reso to help me and other users discover patterns over time—without manual tags or folder maintenance.
The Core Idea
Every transcription gets converted into a semantic representation (an embedding). Notes with similar embeddings are related.
If enough related notes accumulate, they surface as a "theme" with an auto-generated name.
Example: You record five different thoughts about "API design" over two weeks. Reso groups them automatically and labels the theme "API Design Patterns."
The Clustering Challenge
The algorithm itself is straightforward. But tuning it felt impossible:
- Too aggressive: Everything becomes one giant theme ("Work Stuff")
- Too conservative: Every note is its own theme (useless)
I spent days tweaking similarity thresholds before realizing: There's no perfect threshold.
What worked: Adaptive clustering.
Instead of a fixed threshold, we look at the distribution of similarity scores. If there's a natural gap (e.g., scores are either >0.75 or <0.3), we use that gap as the cutoff.
This means the "right" threshold emerges from the data, not from my arbitrary tuning.
Naming Themes
This was surprisingly hard. How do you generate a theme name from note contents?
I tried:
1. LLM summaries: Accurate but slow and expensive
2. Keyword extraction: Fast but generic ("Project Update #47")
3. First note title: Simple but often misleading
What I settled on: Hybrid approach.
We extract frequent keywords, then use a lightweight LLM call to combine them into a coherent phrase. Cost: ~$0.001 per theme. Fast enough to run on-demand.
What I Learned
The best AI features don't feel like AI. They feel like the app is just... paying attention.
When a theme surfaces, people don't think "wow, cool clustering algorithm." They think "huh, I have been thinking about that a lot."
That's the magic—making the invisible visible.