3 ms·

> Personal Library Science is the leverage of LLM technology, applied to a personal library. A personal library differs from a impersonal library in the fact th

> Personal Library Science is the leverage of LLM technology, applied to a personal library. A personal library differs from a impersonal library in the fact that a personal library is an interpretation of a source material. These interpretations include: photographs from different photographers at the same event, or favorite scenes from a movie, or favorite passages from books, parts of songs that bring you to tears, etc. Importantly, these interpretations create unique sets that go on to create unique problems which require unique, idiosyncratic solutions. Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item, or could it derive personal interpretations from user history and the uniqueness of curated items/sets? For those who have been using local, offline LLMs with a manually curated text/image corpus, what have been the most valuable or surprising use cases? Author demo video (2023), https://youtube.com/watch?v=7TgqMRz2r3M https://youtube.com/watch?v=7TgqMRz2r3M & tooling comment (2024), https://news.ycombinator.com/item?id=39789712 https://news.ycombinator.com/item?id=39789712 > Inspired by the commonplace book format, I take highlights from Kindle and embed them in a DB. From there I build (multiple) downstream apps but the central one, Commonplace Bot is a bot that serves as a retrieval and transformer for said highlights. Related: https://en.wikipedia.org/wiki/Lifelog https://en.wikipedia.org/wiki/Lifelog

dartos 3y ago
> Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item No. In something like this you’d probably have the LLM annotate and curate your personal library for you. Potentially by creating and assigning tags or topics based on the content of your library.
- unshavedyak 3y ago
  Yea. I had this discussion not too long ago about this. I'd love to have a combination of a library (Personal Knowledge Management style), data ingestions, and a current world view/state. The PKM is the stored info to write to and query against (both for LLMs and humans). The data ingests are just a pipeline of digital inputs to the system, like chat logs, maybe (transcribed) webcam feeds, files i'm currently editing on desktop, browsing history, etc. The current world view is the interpretation of what i'm doing - to tie all the ingests together and give them context. Eg in isolation browsing some Rust crates might not be that useful. But if i'm also editing Project X on my computer then it's reasonable to assume the searching is related to X. However if it's been 8 hours since any Project X activity, it's less likely related. Same goes for context-less chat logs (as happens frequently in my house) where they are extensions of a voice conversation, etc. All of this stuff is of course insanely privacy invading, so i'd only implement this locally. I also wouldn't even store most of it for fear of data invasion, but using it to fuel a PKM automatically seems pretty sexy. Like browser history, but for your life. This is all just wishful thinking though, LLMs have been moving too fast for me to even bother toying with this. I should note though that i did not intend for LLMs to be "smart". Rather, in a RAG-like fashion (i think is the term), i want to just let LLMs do what they're good at - summarization & autocomplete, and let the world view / PKM store the real data.
  - dartos 3y ago
    FWIW LLMs have been advancing on benchmarks but the practical usage of them (RAG, React, CoT, etc) hasn’t really changed much in the past year.
_bramses 3y ago
> Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item, or could it derive personal interpretations from user history and the uniqueness of curated items/sets? I’ve personally found that tagging is less robust than LLM embeddings (mainly due to dimensionality), but human appended thoughts about a source — also embedded — serve even better as tags. Example: “this is a quote about dinosaurs…” (Old way of doing things) Tags: dinosaurs, jurassic, history Query: “dinosaurs” > results = 1… (New way of doing things) Embedded Quote: [0.182…] User Added Thought: “this dinosaur reminds me of a time i went to six flags with my cousins and…” Embedded User Added Thought: [0.284…] Query: “dinosaurs” > results = 2 (indexes = sources, thoughts) The "thoughts" index can do a second layer cosine similarity search and serve as a tag on its own to fetch similar concepts. Basically a tree search created by similarity from user input/feedback loops.
skybrian 3y ago
I imagine an LLM could work well for doing autocomplete while saving and annotating documents. But it’s not personal unless you edit the result to say what you want to say.

dartos 3y ago

> Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item No. In something like this you’d probably have the LLM annotate and curate your personal library for you. Potentially by creating and assigning tags or topics based on the content of your library.

unshavedyak 3y ago
Yea. I had this discussion not too long ago about this. I'd love to have a combination of a library (Personal Knowledge Management style), data ingestions, and a current world view/state. The PKM is the stored info to write to and query against (both for LLMs and humans). The data ingests are just a pipeline of digital inputs to the system, like chat logs, maybe (transcribed) webcam feeds, files i'm currently editing on desktop, browsing history, etc. The current world view is the interpretation of what i'm doing - to tie all the ingests together and give them context. Eg in isolation browsing some Rust crates might not be that useful. But if i'm also editing Project X on my computer then it's reasonable to assume the searching is related to X. However if it's been 8 hours since any Project X activity, it's less likely related. Same goes for context-less chat logs (as happens frequently in my house) where they are extensions of a voice conversation, etc. All of this stuff is of course insanely privacy invading, so i'd only implement this locally. I also wouldn't even store most of it for fear of data invasion, but using it to fuel a PKM automatically seems pretty sexy. Like browser history, but for your life. This is all just wishful thinking though, LLMs have been moving too fast for me to even bother toying with this. I should note though that i did not intend for LLMs to be "smart". Rather, in a RAG-like fashion (i think is the term), i want to just let LLMs do what they're good at - summarization & autocomplete, and let the world view / PKM store the real data.
- dartos 3y ago
  FWIW LLMs have been advancing on benchmarks but the practical usage of them (RAG, React, CoT, etc) hasn’t really changed much in the past year.

_bramses 3y ago

> Would an LLM-driven "Personal Library" require manually annotated textual interpretation of each curated item, or could it derive personal interpretations from user history and the uniqueness of curated items/sets? I’ve personally found that tagging is less robust than LLM embeddings (mainly due to dimensionality), but human appended thoughts about a source — also embedded — serve even better as tags. Example: “this is a quote about dinosaurs…” (Old way of doing things) Tags: dinosaurs, jurassic, history Query: “dinosaurs” > results = 1… (New way of doing things) Embedded Quote: [0.182…] User Added Thought: “this dinosaur reminds me of a time i went to six flags with my cousins and…” Embedded User Added Thought: [0.284…] Query: “dinosaurs” > results = 2 (indexes = sources, thoughts) The "thoughts" index can do a second layer cosine similarity search and serve as a tag on its own to fetch similar concepts. Basically a tree search created by similarity from user input/feedback loops.

skybrian 3y ago

I imagine an LLM could work well for doing autocomplete while saving and annotating documents. But it’s not personal unless you edit the result to say what you want to say.