Can a local AI do a literature review?
The reading and the cross-checking, yes. The argument, no. Here is what a local tool actually contributes to reviewing literature, and where the line sits.
Yes, a local AI can do most of a literature review's mechanical work: hold every source at once, answer questions across all of them, verify each claim against the passage it cites, and keep notes beside the evidence. What it cannot do is the review itself. A literature review is an argument about how sources relate, and that judgment is yours. The tool's job is to make sure nothing in the pile gets lost, and that nothing gets asserted that the pile does not support.
The parts a local AI is good at.
A literature review starts as an ingestion problem. Papers arrive as PDFs, sometimes a dozen formats deep, and half the work of a review is simply knowing what is in the pile. A local tool turns that pile into a searchable library: every document is extracted, deduplicated and embedded, so a question reaches all of it at once. You stop remembering which paper said the thing and start asking what the pile says.
The cross-source question is the other mechanical win. "Which of these studies report the effect under blinding?" is a retrieval question, and a grounded tool answers it by pulling the passages that bear on it, from every source that has them. Doing that by hand means a day with browser tabs and a highlighter.
There is a privacy reason to want this local, too. A thesis before submission, a grant application, a review of unpublished clinical work: none of it should be uploaded to a service to be processed. On a local tool the sources never leave the machine, and on hardware as modest as a GTX 1650 the models run fine.
The line the tool cannot cross.
A literature review is not a summary of sources. It is a claim about a field: what is established, what is contested, what nobody has looked at. No model makes that argument, and one that tried would produce the worst kind of review, fluent and unaccountable. The tool reads and retrieves; you decide what the state of the field is.
What a well-built local tool does instead is keep you honest while you work. Istor's grounding gate checks every claim in an answer against the passage it cites before you see it, and when the library does not contain an answer, it refuses and says what a real answer would need. That refusal matters in review work, because the failure mode of a confident model against a thin pile is invented consensus.
Notes are the third piece. A review accumulates findings over weeks, and they are only useful if they stay attached to their evidence. In Istor notes live in the same window as the sources they came from, and every answer already carries citations back to the passages it drew on.
How the loop maps to review work.
Import the pile, in whatever formats you have. Ask the cross-source questions, and read the cited passages rather than trusting the prose. When the answer names a gap, that is usually the review's contribution taking shape. Pull web sources only when you choose to, and note what you find beside the answers. When the tool refuses, believe it: the library is telling you the claim is not in there yet, which is either a missing source or a finding.