Autoflow, a Graph RAG based and conversational knowledge base tool | Hacker News Reader

Autoflow, a Graph RAG based and conversational knowledge base tool

github.com

280 pointsjinqueeny36 comments

I ask "what is TiDB" in the demo as suggested, and it takes 2 minutes to start responding in the midst of a multi-stage workflow with several steps each of graph retrieval, vector search, generation, and response combination.

Each of these is individually cool, but it strikes me as tragic that so much effort has been put into an intricate workflow and beautifully crafted UI only to culminate in a completely useless hello-world example, which after 5+ minutes of successive querying and response-building concludes with a network error.

I could use this to build exactly what I need...after stripping out 80% of the features to make it streamlined and responsive.

Why isn't that minimal version the default?

What would you remove?

It appears to be much faster on more specific questions (like the ones that are suggested after you ask it "what is TiDB"). I got a response in about 40s on the question "How does TiDB's cloud-native design enhance its scalability and reliability compared to traditional MySQL databases?"

Also, what's wrong with a nice UI? It appears to mostly be components from https://ui.shadcn.com/. Is there something wrong with good frontend craft, especially for a demo where you're trying to sell something?

It seems like something that is being offered as a self-contained tool that's easy for end users to play with, which isn't going to be the minimal version. I'm sure you could build something that suits your needs exactly, but it would be hard for someone else to predict your exact needs, and there's a decent chance everyone needs or wants a slightly different set of features, and that those things may not make for the most ideal demo.

I am personally far from the typical profile of an AI booster, but I can't help but say something about what I feel is a middlebrow dismissal.

Is this wholly self-hostable? I'd be curious to run something like this on a home server, have some small model via ollama slowly chew through my documents / conversations / receipts / .... and provide a chat-like search engine over the whole mess.

Here is how I am implementing something close to what you mentioned. In my setup, I make sure to create a readme.md at the root of every folder which is a document for me as well as LLM that tells me what is inside the folder and how it is relevant to my life or project. kind of a drunken brain dump for the folder .

I have a cron job that executes every night and iterates through my filesystem looking for changes since the last time it ran. If it finds new files or changes, it creates embeddings and stores them in Milvus.

The chat with LLM using Embeddings if not that great yet. To be fair,I have not yet tried to implement the GraphRAG or Claude's contexual RAG approaches. I have a lot of code in different programming languages, text documents, bills pdf, images. Not sure if one RAG can handle it all.

I am using AWS Bedrock APIs for LLama and Claude and locally hosted Milvus

Wondering if you have tried AnythingLLM, and if so what you thought of it.

I have not .. but this seems to be something I must try.

Thanks a lot, this is the first time i saw a RAG using DSPy. I wanted to know about the expected cost. A few days ago fast graphrag compared their implementation with Microsoft:

> Using The Wizard of Oz, fast-graphrag costs $0.08 vs. graphrag $0.48 — a 6x costs saving that further improves with data size and number of insertions.

I'd love to see a GraphRAG browser that collects the pages I visit automatically.

Many years ago there used to be a Firefox extension (..or might have even been a Mozilla one..) that would store all the pages I visit. I recall its name was Breadcrumbs but I could be misremembering. Space is cheap, or at least affordable if one would exclude videos, which are probably technically more difficult to archive anyway, but sometimes one remembers having seen content that is never to be found again.

I think it would be useful to have just a personal basic search engine on that kind of contents, but possibly a RAG or even a fine tuned LLM would be even cooler.

Actually, e.g. Firefox could do that at least for its bookmarks and tabs, though it already does provide the function for tagging bookmarks. And I think there's probably an extension for searching tabs' contents..

Given how personal browsing history can be this is a great use case for local LLMs. I would love for Mozilla to deliver on this.

Building personal assistant could be beneficial to Mozilla based on how much we do online. I would like to track changes to my beliefs based on how I came across new information. In future, the AI could automatically shorten paragraphs in essays about topics or terms I am already aware of while keeping new concepts introduced in it full expanded so that I grok them better.

The original version of read it later (now Mozilla owned Pocket) had that option. but then removed that option because it went against their commercial interests.

Pocket is good. I use it across all my devices, simple and works for me but do wonder if they could or should do more with the data they collect from me which is all the things I really care about.

What's the selling point for it though? I don't get it?

I need this so much, someone please build it ASAP. This would be so useful!

Working on it https://ont.fyi

The approach is not to capture all pages you view, rather you can add the pages etc you want in order to reduce the amount of noise/rubbish. It constructs a knowledge graph from these documents, and then a graph rag approach ontop to enable chat.

The core graph is based on wikidata, you can have your graphs either private or public if so they are published like those you can see on the site now.

Lots to do, but making ing good progress, if this sounds like something you might want to use please sign up.

After debating for a few days, I joined the waitlist. My Gmail username is the same as my HN username.

The website does a poor job of conveying what the product is exactly, which is the reason that I did not sign up sooner. You might want to approach the text of the site from the premise that the user has no idea what you are presenting.

Thanks for the feedback, and for joining the list.

I agree with your points, my plan is to get a basic version in front of some people and shape it on their feedback. Plan to have it available in a couple of weeks.

Not identical but I started building a smart bookmark tool that stores the content in vectors and sqlite dB and hosts them in GitHub issues with labels managed by the ai. Check it: https://undecidability.com and code lives at https://github.com/irthomasthomas/label-maker It's a bit rough but there is a working cli. It uses local jina embeddings model but openai logprobs to determine when to create new labels.

According to HN and Reddit that would be spyware and and you are wrong for wanting that.

Only if it’s turned on by default and uploaded to the cloud. Privacy and user choice are what these readers want

That's exactly what Recall is: offline and fully customizable, but HN/Reddit went mad over it.

They got mad because you got Recall in an update, no matter whether you wanted it or not, and after another update you couldn’t uninstall it anymore. No choice.

Recall isn't even released yet.

> offline and fully customizable, but HN/Reddit went mad over it.

...until it isn't.

A self-hosted open-source project you can download and run (or compile yourself and then run) is very different from a closed-source OS-level component that's developed by a for-profit company that makes at least some portion of its revenue on ads.

Twitter was "the public square of the web", until it wasn't. Google Reader was a best-in-class easy RSS reader, until it wasn't.

If you don't have the source code, you don't own or control the software. And when you don't own or control the software, it's reasonable to have more-guarded views on what data you're willing to give to that software.

If that software suddenly appears installed on your machine, constantly recording your screen and running entirely-opaque "AI processing" on it, unless you go through a series of steps to opt out...it's reasonable to be upset, because the opportunity to choose what you're willing to share has been denied to you.

And since it's a closed-source OS component, it's only something you can opt out from....until it isn't.

I’d love to see a brain interface so that all these pages we visit can instantly become available to our own non-ai in-brain all-human reasoning.

Local archiving tool I've been testing: webchiver.com

FYI the 'StackVM' link that pops up appears to show all inbound messages.

https://stackvm-ui.vercel.app/tasks/3710e8d2-fb66-4274-9f78-...

Hi, this link is currently for demo purposes. With the help of StackVM, we can DEBUG a RAG retrieval flow step by step and reevaluate the retrieval plan.

Sure, security expectations for a demo are ~0, but “everyone can see everyone else’s inputs” is surprising even by demo standards

Oh, this looks pretty well made. Since it's using nextjs and shadcn/ui, I wonder if they also used v0 to generate components.

Has anyone any experience with TiDB? Haven't heard about it before this post

Yes I have some experience with TiDB. It is pretty amazing actually. They came up with a novel way of distributing data across nodes and having strong consistency while also maintaining great performance. We are recommending it to some of our clients who are looking for an easy scaling option with MySQL (TiDB is MySQL compatible on the connector level.)

It looked neat but relies on a cloud db called 'TIDB', I checked its repo out and it looks like you can self host that as well but damn - it's a lot of containers. So yeah looks like self hosting is an option but likely a pain in the ass.

Wouldn't that be a challenge for most cloud native db's? Are there cloud db's that are easier to self-host?

I couldn't see anything that made it especially cloud native, it doesn't rely on any cloud services - it's just a bunch of containers.

Reply on news.ycombinator.com