_vxw6

Haha I copied it from that page, hilliarious!

I'll keep it haha

It stores it in browser local storage using IndexDB, if you have access to years of documents, it might take up to 10+gb of persistant storage.

gb's of storage, potentially 10+ for very large datasets.

Minutes, not days. Very big data sets might take 30+ minutes (or even a couple of hours), but usefulness starts in the first few minutes (because of the priority algorithm)

Actually a few hundred documents is really no biggy, my current benchmarks is in the range of <250ms (instant feeling) for hundreds of thousands of paragraphs.

I'm testing this on a large knowledge base.

Same bundle of weights but is being run by some rust code that is compiled down to WebAssembly.

It will be very soon https://github.com/haystackoss/haystack

some rust code that compiles to WASM loads LLM from memory, and uses custom transformer.py like rust alternative we wrote.

Forgive me if I don't understand, but I don't think there's a problem with multiple companies using the same common noun as the base for their domain.

Let the best product be remembered for the name.

Not really, gethaystack searches your tabs

Sorry didn’t assume otherwise, replied accidentally to the wrong thread

Note taken lol wow

Thanks

I’m planning on releasing an open source version of this.

But also have paid features that managers would like to use.

I’m using a fine-tuned t5-small model, I fune-tuned it for two tasks, question answering from a paragraph, and highlighting relevant text of search results.

The tricky part is to understand which parts are you indexing, paragraph vs. sentences vs. pages.

But yeah

I load the 90mb model into memory from IndexDB, and then some rust code compiled down to WASM does calculations with the model.

Hi, didn’t intend to repost this, it’s just that HN is hard to figure out.

And most first posts don’t get the intended traction for various reasons (too long, bad wording, unclear).

reposting and changing the post is totally allowed :)

open & free for self hosted version, current alpha version is client side (runs in the browser).

You hit the nail on it’s head, some of these is something I’m dealing with right now (i.e dates)

Makes me wonder, what kind of information sources do you use at work?

Slack, teams, confluence or notion? airtable? jira?

Very good explaination!

hundred of millis is what I experienced.

Actually, the page rank is really based on semantic similarity and relevance of the matched paragraph.

Which under the hood is based of a t5 encoder

Not a designer, I appreciate this advice immensely, I’ll try!

In the setup process you sign in via SSO to all integrations, the token is saved in local storage. That token is used for indexing, and so if you don’t have access to info, the index doesn’t have access.

That’s extremely interesting, I would argue that the reason for keeping tabs open varies, but is something along the lines of: re-reaching the page in the tab is too slow

Yes it needs to stay open. if that’s a problem, I thought of building an extension for continuous indexing.

I’m adding some technical details!

Haystack runs entirely client sided in the browser, so it has a unique tech-stack:

Storage

using IndexDB, haystack stores user indexes locally, + a compressed 90mb NLP model (t5-small) is stored.

Indexing

Locally in the browser, using a t5-small bi-encoder, and some parsing of documents happens in wasm.

Search

Query converted to embedding, then searched over index, atlast results are reranked with a t5-small based cross encoder, and top results go through a seq-to-seq transformer to produce a nice consise textual answer.

you forgot the most important one: haystack.it!

I would argue that every known noun are the first domains to get registered. My goal is to associate workplace search engines with haystack.

What was the name of your product hmu hey@haystack.it, would love to learn from your experience.

It's self-hostable, doesn't store your data, more focused on developers and bottom up motions, I want you to be up and running within 10 seconds.

This is heart warming to hear! Happy new year!

lol ok, come with your buddies ;)