its_down_again

I recently had some luck turning an excel tracker that lists multiple locations and their services into markdown for RAG. It worked great as a natural language lookup, way better than digging through a big Excel sheet.

I uploaded them through Supabase Embeddings Generator if you're curious. https://github.com/supabase/embeddings-generator

But things got a bit messy when I handed it off to someone else. They started using synonyms for locations, like abbreviated addresses to refer to certain columns, which didn't return the right documents.

Followed a friend's suggestion to try NotebookLM, so I uploaded the same docs there, and it was awesome. Some cloud-hosted vector DB tools only handle PDFs, but NotebookLM accepted my Markdown and chunked the docs better than the Supabase library I was using. It just "worked".

I would swap over to NotebookLM because their document chunking and RAG performance is working for my use case, but they just don’t offer an API yet.

I also gave Gemini a shot using this guide, but didn’t get the results I was hoping for. https://codelabs.developers.google.com/multimodal-rag-gemini...

Am I overhyping NotebookLM? I’d love to know to get on-par document chunking, because that seems to deliver fantastic RAG right out of the box. I’m planning to try some other suggestions I’ve seen here, but any insights into how NotebookLM does its magic would be super helpful.

From what I’m getting out of the paper, it’s more like 'students who work out regularly are less likely to get really bad hangovers.' I’m not sure I buy that regular exercise actually lowers the pain of a hangover if you’re drinking the same amount. For me, working out tends to make me feel tipsy faster, so I usually end up drinking less anyway when I do.

Living in San Francisco, I feel much safer as both a pedestrian and a cyclist when I see a Waymo at crosswalks vs. human drivers. I was hit by a car once while cycling because the driver zoned out, realized too late they were about to miss a turn, and struck me while attempting a last-minute turn. Since then, I’ve become hyper aware whenever drivers are on their phones at stop signs or play pointless games of chicken, trying to beat red lights just to end up stuck blocking the middle of the street.

When I see a Waymo, it feels like it's aware of me and always stops at a safe distance, and I don't fear a collision. It doesn't give off the hostile or dangerously clueless vibe I often get from human drivers in the city.

I’ve spent some time in enterprise TFO/demo engineering, and this kind of generative tool would’ve been a game changer. When it comes to synthetic data, the challenge lies at the sweet spot of being both "super tough" and in high business need. When you're working with customer data, it’s pretty risky—just anonymizing PII doesn’t cut it. You’ve got to create data that’s far enough removed from the original to really stay in the clear. But even if you can do it once, AI tools often need thousands of data rows to make the demo worthwhile. Without that volume, the visualizations fall flat, and the demo doesn’t have any impact.

I found challenge with LLMs isn’t generating a "real enough" data point—that’s doable. It’s about, "How do I load this in?", then, "How do I generate hundreds of these?" And even beyond that, "How do I make these pseudo-random in a way that tells a coherent story with the graphs?" It always feels like you’re right on the edge, but getting it to work reliably in the way you need is harder than it looks.

I don't have a daughter yet, so can't really comment on that perspective. But I've had a great experience with the SF public libraries. I visit the main one near Civic Center pretty often just to use the printer, and honestly I just love being around stacks of books. I haven't been since the pandemic, but the one in Chinatown was always packed with kids studying together, and it was clean.

I grew up going to libraries in the suburbs of Chicago and Philadelphia. Those are some of my best memories. I wouldn't have developed the computer skills or built the career I have today if I hadn't seen other kids, a little older than me, using them. Just being able to spend hours in a quiet, safe space where people are learning for fun— I think it changed my life.

It's always situational. If you live in a city and want to get away from people for 40minutes+ on the daily, it's not that practical to drive far away until there's no traffic. If you live in San Francisco, it's relatively easy to reach low traffic roads on bike in the morning, or find less crowded paths by running there.

Of course if you're rich and and feel like it, you can get a nice crashpad in Napa for weekends and drive your Koenigsegg downtown for a morning coffee.

From https://www.whatgenerativeai.com/docs/genai-playbook/unveili...

"Many GenAI models use unsupervised learning, allowing them to find patterns in data without explicit labeling. This helps models generalize beyond their training data and create new outputs."

I'm curious how unsupervised learning relates to the transformer architecture or RLHF. What aspects of these methods, or other GenAI technologies you know of tie into unsupervised learning?

Honestly, it looks really great. Definitely more polished than what I was expecting from open-source/hacker setup.

I've been dailying a Breville dual boiler since 2019, and I've tried a few other machines at coworking spaces and friends' places over the years. For me, as long as I can hit 9 bars with a decent puck extraction, the machine itself doesn’t make a huge difference in flavor. The challenge is finding the right beans, and the fine-tuning through the grinder and puck prep.

It's not that my machine struggles to push water through consistently, it's that I struggle to nail the grind size, achieve even particle distribution, and tamp evenly. Mornings especially—I just don’t have steady hands right after waking up. Funny enough, my best shots are usually in the afternoon. Sadly I try to avoid drinking coffee past noon, so it’s often for someone else, and I just take a tax with a tasting spoon.

The way I see it, upgrading the machine helps with drinks per hour throughput, not single shot quality. I think that's also why the at-home coffee community is more into things like the niche grinder and blind shaker.

That's awesome! Mind sharing the prompt you used? I'm working on something pretty similar but a little more complex—looking to convert from an Excel sheet to Markdown or JSON.

Hey it's working great! Small nitpicks while I'm testing on my macbook + chrome. - It can be easy to "miss" the next card because scrolling feels really fast - I accidentally uploaded the same files twice, but can't figure out how to delete the duplicates.

Nice work! I saw that the prompt will generate 5 different queries to test, but I'm not sure if metadata is included/needed in this project. Let's say I have an encyclopedia of birds with sections that look similar, like "Species Name: {NAME}" or "First Discovered: {Date}", and I upload all the pages. If I search for something like "peregrine falcon habitat" or "hummingbird diet," will it give me the right results?

Totally dig the idea! Can you throw in some sample fake company docs that have worked well before to test out the demo?

Thanks! Just trying to get a clearer picture—what websites or products were you keeping an eye on before? Got something specific you're looking to snag right now?

Thanks for sharing Visualping— can't believe I hadn’t heard of them yet! They seem to be class-leading in “web change detection notifications”, even serving some F500 clients. I checked out some third-party reviews and it seems they sometimes gets complaints around reliability and pricing. That might explain why people have spun up open-source alternatives like changedetection and Huginn, which are more powerful and flexible, but don't match Visualping's user-friendliness and customer support.

I’m definitely doubling down on the agency use case since I have a connection there. But based on what I’ve gathered, here are a few other potential uses:

- For tech marketing managers managing multiple products daily, ensuring everything’s purchase-ready. - For developers, monitoring forums for important updates, filtering out the noise. - For warehouse managers, keeping an eye on product pages for restocking purposes, maybe checking in about once an hour depending on demand. - For marketers, detecting website outages—this should run as often as possible without causing issues. - For academics, staying updated on relevant research grants to quickly apply. - For me personally, as a marathon runner who switches out training shoes every 300 miles or so, I want to catch discounts on last season’s models which perform just as well at half the price. - For everyday consumers, checking international shipping updates, which can sometimes require multiple form submissions.

It also appears perception of web clipping tools is "advanced mode" and not easy to use. I’m excited about making them simple enough for anyone to use. I'm imagining an interface where you can easily save what matters to you, record your activities, and keep those sessions safe and private on your own device for future replay. Also hoping that making it open source builds trust, but I'm still checking how well that works. The goal is a tool that "just works" out of the box while still keeping your data local, private, and secure.

Also noticing that notification filter setup can be complex, and require a lot of upkeep. I’m thinking about how to make that easier and more reliable. Currently plan to use a fine-tuned local LLM for parsing & interpreting HTML and JSON. This approach would help the AI be smarter about timing & sending notifications. For example it would ignore minor changes like a one-minute increase in your commute, but alert you about significant changes like a sudden 15-minute jump.

Let me know if you have any more thoughts on this! Also if you’re open to it, I’d be interested in learning more about how you use Visualping.

Hey, I like what you're doing with duckboard! I checked it out and it's actually quite similar to what I'm working on.

I was curious about integrating with the "webclips" feature since it was native, but it turns out that was part of the "Dashboard" feature, which got dropped after Catalina. But I did find a Reddit thread where someone had a clever workaround.

Quote from zonkbonkbadonk on Reddit: "It's not the same, but the closest I've found is creating several desktops and using Chrome's More Tools > Create Shortcut in a New Window option to make a borderless, resizable website window. Then, I arrange them on that desktop. I also use Adblock Pro or uBlock Origin to 'hide elements' and clear away parts of the website I don’t want to see."

Here's the link to that discussion if you want to check it out: https://www.reddit.com/r/MacOS/comments/kojjzn/comment/ghrm5...

Hey there! Thanks a bunch for your input—it really helps to think through where this project is headed. Right now, I'm still in between a couple ideas on how to shape the project.

One thing I'm personally keen on is using it to peek at webapps like Teams or Slack without falling down the rabbit hole of distractions. e.g responding to a message without getting sidetracked by unread notifications. But to make that work right, I’d need to handle logins and authenticated sessions, which I know from experience is hard to get right.

Another scenario I’m looking at is for stuff like MX validator tools. When you tweak MX server settings, it takes a good day or two to take effect globally, and it’s crucial not to send out new emails until those new configuration changes are fully propagated.

I’ve also heard from a friend who works at an Amazon agency. They’re constantly refreshing product pages they manage waiting for updates to show up. They’ve tried existing tools, but reliability is often a toss-up. But even at $50 a seat, it feels like a bargain to them!

You're totally right about detecting changes-- making the app capable of reliably spotting updates and notifying users could really be key. I’m thinking about something like to a git commit history for the web. What's your opinion on a viable UI/UX for the notifications? Maybe a desktop notification that shows the new change?

Also really appreciate the tip about using GitHub releases for distribution! I was really overthinking it, but I'll get on that ASAP.

Let me know what you think, I'd love to hear more about the use case if the app could detect changes and signal the user from the background.

Hey good call out! What version of Mac are you using? I just packaged up a .ipa file that you can install directly to your applications. It’s up & running OK on my M1 air w/ Sonoma 14.5.

Here’s a link where you can grab that file if you'd like to give it a try: https://drive.google.com/drive/folders/1Kt6fdFhxdGQFVbmOWRAV...

Let me know how it goes!

Thanks for the link! That looks pretty similar to what we're doing with Glanceables, and I like how it ties right into the macOS dashboard. Have you used it yourself? I’d love to hear your experience with it!

What’s your last few post heavy night 5K times?