How to chat with your bookmarks
To chat with your bookmarks you need a tool that reads what you saved rather than the web, so retrieval runs over your own items and the answer cites them. In Kyurations this is Ask Kyuri, and it is free. Save enough that there is a corpus to search, ask by meaning rather than by keyword, check the cited cards before you trust a claim, narrow to a single list when an answer comes back too broad, and turn a good answer into a list so the next person does not have to ask again.
The steps
- 01
Save enough to have something to ask
A library of twenty links has nothing to tell you. The useful threshold is roughly the point where you can no longer remember everything in it, which for most people is a few hundred items.
- 02
Ask by meaning, not by keyword
Describe the argument rather than guessing the title. Retrieval runs over what a piece is about, so 'the video that argued taste comes from proximity, not training' finds the thing whose title you never knew.
- 03
Read the citations before you trust the answer
Every claim should open the card it came from. If a sentence has no card behind it, treat it as the model talking rather than your library, and check it.
- 04
Narrow to one list when the answer is too broad
Scope the question to a single list and the same query gets sharper, because the corpus it searches got smaller and more coherent.
- 05
Turn the answer into something you keep
Ask it to make a list out of what it just found, or to tag those items. Otherwise you will run the same query again in three months.
Why a general chatbot cannot do this
Ask ChatGPT what you saved about a subject and it will tell you about the subject. It has never seen your library. What comes back is the average of what has been written, which is exactly the thing you were trying to escape when you started saving the good ones.
The value in a collection is the judgement in it. Every save is a small decision that this was worth keeping, and a few hundred of those decisions is a filter nobody else has. An assistant answering out of that corpus is answering out of your taste. An assistant answering out of the open web is answering out of everyone's.
What good questions look like
- Recall: the piece you remember the shape of but not the name of.
- Synthesis: what four saved articles actually disagree about.
- Stock-taking: how much you have saved on a subject, and whether it is enough to write from.
- Resurfacing: what you saved a year ago on something you are thinking about today.
What it will not do
It cannot answer from things you never saved. That sounds obvious and is the single most common disappointment: a library is a small corpus, and a question outside it should come back as 'nothing here covers that' rather than as a confident paragraph.
It is not a search engine with better manners either. If you know the exact title, search is faster and always will be. Chat earns its place on the questions where you cannot name what you are looking for.
Questions people ask
- Can I chat with my bookmarks?
- Yes, if the tool holding them reads them. Kyurations does this through Ask Kyuri: it searches your saved items by meaning, answers from them, and cites the ones it used so you can open each source. It is free and needs no set-up.
- How is this different from asking ChatGPT the same question?
- ChatGPT answers from what it was trained on. Ask Kyuri answers from what you decided was worth keeping, and shows which of your items each part of the answer came from. On a question about your own collection, the small curated corpus wins.
- How many bookmarks do I need before this is useful?
- Enough that you cannot hold them all in your head, which is usually a few hundred. Below that you can still scroll the list faster than you can ask a question about it.
- Can it organise my library as well as answer questions?
- Yes. It can create a list from what an answer found, add or remove items, tag them, save a link or a note, and publish a list. Deletes are a recoverable soft delete, and every write is one you asked for in the thread.