Status: concept study. This is a proposed experiment, not a working product or a report of measured results.
The question is deliberately narrow: can a small assistant help someone find and understand information in a folder of notes, while keeping the underlying documents on their own computer?
Define the boundary
The first version would accept a selected folder of Markdown and text files. No email integration, no background synchronization, and no automatic access to the rest of the machine.
Documents would be split into passages and indexed locally. A query would retrieve a small number of relevant passages. A locally running model would use those passages to draft an answer with references to the source files.
“Local-first” is a design intention, not a privacy guarantee by itself. Each dependency, model download, logging setting, and network request would need inspection. The experiment should make those boundaries observable.
The proposed path
- Select a small set of non-sensitive notes.
- Build a local index with a record of each source path.
- Retrieve candidate passages for a question.
- Ask a local model for an answer grounded in those passages.
- Show the passages beside the answer so the reader can verify it.
A source reference is not proof that the answer is supported. The interface should make it easy to compare the claim with the actual text.
What I would test
The first evaluation would use questions with known answers, questions that require more than one note, and questions the collection cannot answer.
I would record whether the system finds the right passage, whether its answer stays within the evidence, and whether it admits missing information. I would also compare it with ordinary keyword search. If the simpler tool works better, that is a useful result.
Open questions
How often should the index refresh? What happens when a note is deleted? Does the model treat instructions inside a document as data, or try to follow them? How much memory does the entire workflow need on an everyday computer?
The aim is not to build a universal assistant. It is to learn what a small, understandable one can do well.