Home/The library
Research · Los Cabos
A library with the source in hand.
We built a research system over a business podcast. Every episode we analysed was saved as a note with its link, who said it, and when. When you ask something, the answer arrives with the citation — not with a summary nobody can check.
Analysis of the archive is paused on purpose. We would rather tell you exactly what is missing than sell you a complete library that does not exist.
- Scoping a product down
- When to stop pushing
- How to organise the work
- What was said about pricing and customers
The cut-off, in plain sight
These numbers are the product, not the small print.
Almost any system can produce an answer. The hard part — and the part almost nobody shows — is saying precisely where an answer came from and where the knowledge ends. So the cut-off goes first:
372 episodes
- 272 with saved, searchable notes
- 99 still to be analysed
- 1 with no available transcript
The 99 pending are not counted as covered. If your question lives in one of them, the system says so instead of filling the gap.
1,735 vectors
- 1,024 dimensions, cosine similarity
- BGE-M3 embeddings
- Search in Spanish and English
Every result resolves back to the full note: episode link, speaker, date, time anchor, caveats and the hash of the source text.
12 source episodes
- A provisional synthesis
- Coverage kept separate from the 272 notes
The methods profile is the part most used when applying an idea, and it is the narrowest thing we have. It is labelled provisional and was never quietly promoted to a full-archive synthesis.
Paused
- No new episode crawling
- No background analysis running
- No promise of full coverage
We would rather have an honest library that is stopped than one that keeps growing while it stops being able to explain itself.
How it answers
It searches, it cites, and it tells you how far it reaches.
A business question
Specific, in Spanish or English. You do not need to know the episode or who said it.
Across the saved notes
It finds the passages that address the topic, and reports the coverage: how many episodes answer that question.
Name, date and link
Every passage keeps who spoke, when, the episode link, and the caveats that came with the idea.
And what it would take
If the question needs an episode that is not analysed yet, it says so. That gap is information, not an error to hide.
The limits
What the library does not do.
Written here because the limits are part of the deal. A system that knows how much it does not know is more useful than one that always has an answer.
Why the cut-off is the feature
It is the same discipline we use in the agent.
An agent answering your customers has to know when it does not know. If it invents a price or an opening time, it puts you in trouble. This library was built under the same rules: it answers only from the source, cites where it came from, and when it cannot reach, it hands over to a person.
That is why we show the cut-off instead of hiding it. A system that declares its coverage can be used to decide. One that inflates it cannot.
Would a library that cites be useful to you?
If you have an archive of your own — internal calls, recorded meetings, manuals, support email — we can build the same thing over your material, with your rules and your limits in plain sight.