Local AI for field research.
Fieldwork happens where the internet is thin and the material is sensitive. A research tool that needs neither is built for exactly that.
Field research is the case where local stops being a preference. Interviews, observations and participant data are usually bound by ethics approvals that forbid uploading them to a third-party service, and field sites rarely offer the connection a cloud tool assumes. A local research notebook works the whole loop offline: import the transcripts and notes as they accumulate, ask questions across everything gathered so far, and carry citations back to the exact passage. The one thing it does not do is collect the data for you. The recorder, the notebook and the camera are still yours.
Why the field is different.
Two constraints shape analysis away from a desk. The first is coverage. Remote sites, archive basements, vessels and clinics are full of dead zones, and a tool that needs the network to answer a question is a tool that stops working at the worst moment. The second is consent. Human-subject research usually restricts where participant data may be processed, and "uploaded to a cloud provider" is the one answer an ethics board cannot accept.
Both constraints dissolve when the models run on the machine. Istor's retrieval, prompting and inference are all local. Import a transcript on a train, question it in a hotel with no wifi, and nothing has left the laptop. The only network use in the app is a web fetch you trigger, and in the field you simply do not.
How the loop carries fieldwork.
Field data arrives in pieces. Day one is a recording and a photo of a consent form; day five is four transcripts and a scan of a handwritten log. Istor takes PDFs, DOCX, EPUB, HTML, Markdown and plain text, and its vision model reads scanned pages, so the pile can keep its original shape instead of waiting for a cleanup pass. Each import is deduplicated, so re-adding an updated transcript does not fork your library.
The analysis questions start early, and they should. "Every mention of the relocation" across twelve transcripts is a retrieval question, and the answer arrives with citations to the exact passages, session by session. That traceability is what turns a model's summary into evidence: a claim you cannot follow back to the moment it came from does not belong in a findings chapter.
Notes stay in the same window as the sources they came from, which is where fieldwork analysis actually happens, halfway between reading and writing. When the tool refuses an answer because the sources do not contain one, that is useful too: it is the difference between a theme nobody mentioned and a theme you misremembered.
The honest limits.
The tool does not record, transcribe audio or take the photos. It works on what you bring back, which it reads, searches and questions well. It needs a GPU, though a modest one: a GTX 1650 class card runs the roster, which is the hardware a field laptop already tends to have. And the model on that card is not a frontier model; it is a good local one, which is why every claim carries its citation, so you can check it against the source in a click.