Let's use Ollama's Embeddings to Build an App 

Ollama doesn't really have a concept of session that you can recover. If you are using a GUI frontend for Ollama, some of them may have a way of saving sessions.

That’s a helluva lot of extra cruft for a tiny bit of convenience

Typescript/Javascript tends to get more views. It’s used by far more folks than python as shown by every survey for years. I find it easier to work with and understand. And faster.

@@technovangelist Isn't right that Python is the mostly popular programming language for AI, esp. in learning and research?

Maybe but for no good reason. Python has no benefit here over any other language.

But thankyou for strengthening the argument for doing more JavaScript/typescript content.

I don't know. I watched many AI-related youtube videos. They include school classes and videos from businesses and individuals. Most of the time, they offer implementations in Python or even code working in Colab. If you do a search, you can find out that Python is widely taught in colleges, but not TypeScript/Javascript. Maybe I am wrong. Your video has excellent content, and I feel that I didn't get the most out of it.

This one? Sync Ollama Models with Other Tools ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-UfhXbwA5thQ.html

This video is just about embedding

Docker on Mac doesn’t use the gpu so I’d doesn’t make sense for me to include that. And nowhere near as easy as a straight native install

Oh no, it was your mind I was taking over.....

This...this make so much sense...this explains everything!@@technovangelist​

As one thing? Then stuff wouldn’t come up in a match

​@@technovangelist What you did is very advanced, and probably more effective, but sometimes I wonder if we are not reinventing the wheel, models are able of quite good needle in a haystack with plain text, as long as you remain in a reasonable length of text per promtp. What I did notice is when creating the jasonl it can be inconsistent, but then it takes some trial and error the get the prompt right, or an extra call to repair it, which with Openai becomes quite expensive, so yeah, I am cheating... :) But today I am mourning: bitsandbytes update installed a new triton 2.2.0, that broke flash attention, xformers and, because why not, torch, so I am looking at error logs today, thinking wtf am I doing under the rain wasting my time while I could jump in the car and drive to the south of Spain as they have 35 degrees. But I might isolate the beginning of the flow, for just for dataset creation, it's already a decent tool. It was working quite well before I try to lump half of the finetuning library of Github in my venv. Only I, like you, last time I programmed something, was an online casino in Basic with the 'rnd' function, on a 286. So even the most basic development takes me on a cathartic journey contemplating the abysmal ignorance I bring to the table in terms of development, so I always try to kiss. Peace! 😄 *Edit actually the casino was the opposite of "online" ...

Actually in a lot of cases they are very bad at plucking out the right concepts from a long text. You don't have to spend much time with chatgpt or any of the local models to experience that. But this isn't doing that anyway. This one is all about embeddings which is different from asking a model a question

@@technovangelist Yes I agree, but both solutions, IMO give back 'lumps' of the original text and rag from what I have tested, embedding with various dbs and models is always inferior to finetuning that returns a more organic answer, yet less precise. I guess it's going to depend on the application, probably a mix of both is going to be the best, even if I am only half impressed with OpeanAI's Agents that keep forgetting they have a context. Let us know when you have something we can test, whatever works my friend!

@@technovangelist I've wondered about taking an LLM-generated summary of the raw transcript and using that summary as the embedding, though. That way you get closer-to-deterministic length so it fits nicely in one embedding. With the length of your videos, I'm betting 32K context length would be enough for the transcript + instructions prompt, plus the output (probably even with chain of summaries).

It is. That’s the answer I provide. Normally rag would give the results of a vector db search to the model to generate an answer

for code samples, look at the ones associated with the video...my github is technovangelist and the repo is videoprojects. the model to use is the one in the video...nomic text embed

@@technovangelist found it 🎊🎊