ScrollWise May Update: Stability Fixes and the Upcoming Browser Extension

A Journey of 1,000 Miles

… starts with me forgetting to push to prod for 3 months.

It’s been a while since I’ve dropped an update for ScrollWise. Since the January changelog update, I’ve been hard at work on some updates for ScrollWise that aren’t as exciting as I’d like but are vital to the continuity and usability of the app. Let’s dive into the biggest stability fix before we talk about the exciting things that are coming next.

If you don’t care too much about the nerdy technical breakdown, here’s the TL;DR on what’s new:

• Bugfix for conversation context - There were several bugs in the way that the system was handling newly added resources that made it hard or impossible to add new resources to a topic and have an ongoing chat discover them. These bugs have been fixed and the chat window is much more reliable now
• Stability improvements - There is now a new analytics and monitoring suite in place to find bugs more proactively. The bugs in the conversation context window should have been found earlier and were pretty difficult to track down without proper analytics and monitoring, so this will help a lot in ensuring new changes and bugfixes are made fast.
• Preparations for the web extension - ScrollWise is about to get a huge upgrade in the form of a web extension that brings the power of ScrollWise to every page you visit on the web. I’ve implemented some critical backend changes to make this update easier.

Now on to story time!

---

Bugs Are Gross

A few months ago, a power-user sent me a message that they were having issues with the ScrollWise sidebar chat. Specifically, when they added new content to a topic, the chat wouldn’t “find” the new content. So they were mid-conversation, found a new research paper to add and added it, but the ongoing chat they were having was not able to read the new paper into the conversation.

This isn’t ideal, especially since I’m hoping to add New Content Discovery as a feature to ScrollWise soon. If a chat could only find content that was added before the conversation started, this makes the chat window harder to use, and it makes longer-context conversations more difficult, maybe even impossible.

To explain what was going on, I want to explain a little of the magic behind the ScrollWise system.

When you upload a new resource, it takes the words in that document and “vectorizes” them, turning the words into a huge list of numbers that represent the words. This is the “language” that Retrieval Augmented Generation (RAG) can use to search for relevant sources when you type in chat. When you send a chat message, that chat message is also vectorized, and the ScrollWise system can take your message and use it to search for relevant resources.

When a relevant resource is found, it is added to the conversation’s “context” so that the model can use it in its answer. The relevant text in that resource is sent along with your question to the LLM and the LLM takes the resource, previous messages in the conversation and your most recent question into its context to form an answer.

What was happening previously was a combination of a few bugs. The list of content in a given topic was not being updated, so it wasn’t even able to search for the correct content. That was an easy fix. There was also conditions where the search for relevant content was run before a resource was finished being vectorized, so the new resource couldn’t be found then either since it wasn’t yet in the right format. Finally, the algorithm I wrote to prioritize resources and order them needed to be tweaked so that the model received the right resources.

Now that those bugs are fixed, the conversations you have in the sidebar chat window will get the right context to answer your questions and assist in your research, even if you’ve just added them. These bugfixes are critical for what comes next.

Upcoming: The Web Extension

What’s coming is far more exciting than some bug fixes, and I want to give some sneak previews on what you can expect with the next release.

The ScrollWise Browser Extension is going to bring ScrollWise to every page you visit, natively, to include your favorite AI providers. You will be able to highlight and save content from any web page you visit straight to any ScrollWise topic.

Critically, we’re also bringing the power of ScrollWise to your favorite LLM providers, including Claude, ChatGPT and Gemini, instead of just bringing those providers to ScrollWise. With the web extension, you will be able to seamlessly inject context from the resources you’ve gathered into your conversations with your favorite models on the model providers’ sites themselves. Instead of relying on ScrollWise’s chat window, you will be able to use the native chat functions on your favorite sites much more naturally.

The ScrollWise Vision

This has been the vision of ScrollWise from the beginning. I want to make it as easy as possible for individuals to use AI in a way that empowers their research. Hallucinations will always be a problem, or at least a worry, for AI-enabled research, but being able to inject your own research, from papers on arXiv and esoteric PDFs to YouTube videos and blogs, can add some level of peace-of-mind.

AI has gotten a reputation for enabling lazy writing and slop. People at large, especially outside of tech, don’t associate LLM usage with trustworthy research. Some of this reputation is deserved: AI has indeed enabled a massive influx of short-form video and text content slop, lazy and annoying writing and misinformation. I don’t deny this, even if I use AI on a daily basis for everything from coming up with recipes to writing code. AI has an image problem, one that isn’t helped by a lot of the marketing that AI executives and researchers are taking part in.

I believe that one of the things that can help correct this image problem is for software developers to build and release software that is actively helpful for real, in-depth research, especially if those applications can help alleviate some of the problems that current models have like hallucinations and a lack of access to reliable and up-to-date information.

We as technologists have a responsibility to usher in the new age of artificial intelligence in a way that supports human flourishing. Thus far, we’re doing a pretty lousy job at it. ScrollWise is my humble contribution to reversing that trend.

---

Have questions about the new features? Reach out on X/Twitter or email ceo@radc0rp.com