[scotty79]

Interesting idea ... doesn't fit my style of content consumption, but it might fit yours.

I went for a personal newsfeed, agent pulls news form ~100 feeds related to my interests. Then reads all articles for me and orders them by how interesting they might be for me. I specifically asked for vector embeddings, up/down votes (-2..+2), visited status, LLM content evaluation. Probably there are some other mechanisms I didn't even bother to check. It's a work in progress but I can see myself replacing most of my news reading with it. For many news the AI summary, which contains main idea behind the item is enough for me. As a bonus it resolves clickbait and is quite good at it. Also no ads, ever. For sure I need to implement some grouping because when the popular story breaks I have many stories about the same thing with mostly overlapping details. AI merging them would be quite cool.

I also asked AI to extract my interests from my browsing/watching histories of my all accounts. V2 of my newsfeed might utilize that somehow for better results.

Silly thing is I made it in one afternoon with my only motivation of being slightly more annoyed with the web on that day.

[CamperBob2]

So this actually runs within ollama, which you then leave up full-time? Or only when you manually want to refresh your news feed?

[scotty79]

The prototype runs through ollama on my laptop while I browse. I usually have few hundreds stories ready from the previous session, as new stories come in they get processed in the background while I browse the old ones. It processes stories faster than I consume them.

When I move it to the server I might consider waking it up periodically to pull and analyze new stories and perhaps notify me if something absolutely great shows up.

There are so many possibilities for tuning it. And I don't need to think how to make it secure (beyond the basics), ultra performant, fitting other people's tastes and so on because this program has audience of one.

[CamperBob2]

Very cool, so you just time-shift the good content while removing the rest. This program would have an audience of at least two if it were public.

How do the vector embeddings fit into the picture?

[scotty79]

I was basically throwing stuff at the wall and seeing what sticks and makes a pretty picture for me. Vector embeddings were just one of the things I threw, in hopes of boosting news items thematically similar to the stuff I already liked. I can't really tell how much good or bad are they doing. All I know is that they are fast (embeddinggemma through ollama) and their output is worth few points in some combined heuristics that AI concocted for me.

[mannanj]

Particular model or LLM that you use?

[scotty79]

Local ones (through ollama, probably will use something more performant when I move fully to it).

When I touched it the last time I was a fan of gemma4 https://ollama.com/library/gemma4

Since then I grew really fond of Qwen3.6 (it's super capable in coding against the DOM) so I'll probably try to move to it with the next iteration. https://ollama.com/library/qwen3.6

[mannanj]

Great stuff. Thank you!