The case for an AI that reads your library, not the web
An assistant grounded in your own saved library answers better than a general one on any question about what you have read, because the corpus was filtered by a human with a stake in it. A few hundred deliberate saves encode judgement that no amount of training data reproduces, and grounding an answer in them makes it checkable: every claim opens the item it came from. The trade is coverage. A curated corpus cannot answer what you never saved, and a tool that pretends otherwise has given up the only advantage it had.
A small corpus is the feature
The instinct is that more data is better, and for training a model it is. For answering a question about what you have read it is the opposite. Every item in your library got there because you decided it was worth keeping, and that decision is a signal the open web does not carry.
So the corpus is small, opinionated and yours, and those are the three properties that make the answers feel unlike a chatbot's. It is not smarter. It is better informed about one specific thing, which is what you have been paying attention to.
Grounding is what makes it checkable
An answer with no sources is a claim you have to take on trust, and the failure mode of language models is that a wrong answer looks exactly like a right one. Citations change the shape of the problem: if every sentence opens the card it came from, verifying is a click rather than a research project.
This also constrains what the tool can get away with. When references are index based, pointing at a real retrieved item, a fabricated citation cannot resolve to anything. The honest answer when the library does not cover a question is that it does not.
Where it fails, and should say so
Coverage is the trade. Ask about something outside your library and there is nothing to retrieve, and the right behaviour is to say so rather than to fall back on general knowledge and present it in the same voice as a cited answer. A tool that blurs those two has thrown away the reason to use it.
The other limit is that a library reflects what you saved, including its gaps and its enthusiasms. If you saved eleven articles making the same argument, a summary will report a consensus that exists in your collection and not in the world. Being answered out of your own taste is the point, and it is also the bias.
What this looks like in practice
- Retrieval first, then generation, so the answer is assembled from real items rather than recalled.
- Citations that resolve to cards you can open, not footnote-shaped decoration.
- Full text where it exists, so a question about a long report is answered from the report.
- The ability to act on a result, because an insight you cannot file is one you will look for again.
Questions people ask
- Why not just use a general AI assistant?
- Use one, for most things. The narrow case where a library-grounded assistant wins is any question about what you personally have read and saved, because a general model has never seen your collection and cannot cite it.
- Does answering from my library mean my collection trains a model?
- No. Private collections are not training material. Retrieval means the items are fetched to answer your question in the moment, not absorbed into a model's weights.
- What happens when my library does not cover the question?
- You should be told that. An assistant that quietly switches to general knowledge while still sounding cited is worse than one that has no answer, because you can no longer tell which mode you are in.