Post

Do you want an Obsidian Integration or how you want to trigger it? I set one up locally pulling in from Notion, with a custom frontend, and a Lance vectordb in the backend.

I don't mind the interface; it can be a CLI or a script. No need for a UI, the content and quality of responses matter more. How did your backend "learn" or connect based on your notion pulled data? Does Lance have a custom search/algorithm you use? Or do you add additional steps?

I use hybrid search + a reranker on top of Lance. Did some light relevant engineering and do a weighting based on the different fields (e.g. title, body, headings in the page). I did not do any query understanding because I know how to write my queries to have a good retrieval score.

I am running Ollama with Continue in VSCode, this let’s me ask questions to my “codebase”, which just happens to be the entire vault😃

Ohh that looks great. Thanks for sharing. How hard is it to extend or plug your vault to Continue? And how accurate/useful are the responses?

My move is that I open the vault as the project folder in VSCode, and then Continue has one button to “embed codebase”, takes a few seconds, and then it’s done. I found that it’s good at finding relevant notes, but only 4-6 at a time contd

also interested in this, but haven’t come across anything yet fwiw: github.com/QuivrHQ/quivr or reasonabledeviations.com/2023/02/05/g...

Yes, I have tried it with Flowise as well as building it from scratch using Python, Ollama and LangChain. I loaded the documents to the vector store separately and then used them in my inferencing. If you know the structure of your markdown, you can use it to better guide the chunking process.

I haven't done any additional fine tuning, and used existing fine tuned models.

awesome! Wdyt, to what size or amount of notes does RAG work, and when does "training" a model make more sense? I wonder as I have 7k+ notes with huge notes, that RAG might not be enough, but I'm not experienced in it, so I do not know.

How hard was that? And is the code somewhere on a GitHub or similar? :)

I am sharing the essence of the workflow here: github.com/buzypi/full-... PS: I have not cleaned up the code and there are lot of hard coded values. Hope you can manage with this!