28M Hacker News comments as vector embedding search dataset
clickhouse.com319 points by walterbell 8 hours ago
319 points by walterbell 8 hours ago
Don't use all-MiniLM-L6-v2 for new vector embeddings datasets.
Yes, it's the open-weights embedding model used in all the tutorials and it was the most pragmatic model to use in sentence-transformers when vector stores were in their infancy, but it's old and does not implement the newest advances in architectures and data training pipelines, and it has a low context length of 512 when embedding models can do 2k+ with even more efficient tokenizers.
For open-weights, I would recommend EmbeddingGemma (https://huggingface.co/google/embeddinggemma-300m) instead which has incredible benchmarks and a 2k context window: although it's larger/slower to encode, the payoff is worth it. For a compromise, bge-base-en-v1.5 (https://huggingface.co/BAAI/bge-base-en-v1.5) or nomic-embed-text-v1.5 (https://huggingface.co/nomic-ai/nomic-embed-text-v1.5) are also good.
I am partial to https://huggingface.co/Qwen/Qwen3-Embedding-0.6B nowadays.
Open weights, multilingual, 32k context.
Also matryoshka and the ability to guide matches by using prefix instructions on the query.
I have ~50 million sentences from english project gutenberg novels embedded with this.
What are you using those embeddings for, If you don't mind me asking? I'd love to know more about the workflow and what the prefix instructions are like.
How do the commercial embedding models compare against each other? Eg Cohere vs OpenAI small vs OpenAI large etc?
I have troubles navigating this space as there’s so much choice, and I don’t know exactly how to “benchmark” an embedding model for my use cases.
It's a shame EmbeddingGemma is under the shonky Gemma license. I'll be honest: I don't remember what was shonky about it, but that in itself is a problem because now I have to care about, read and maybe even get legal advice before I build anything interesting on top of it!
(Just took a look and it has the problem that it forbids certain "restricted uses" that are listed in another document which it says it "is hereby incorporated by reference into this Agreement" - in other words Google could at any point in the future decide that the thing you are building is now a restricted use and ban you from continuing to use Gemma.)
One thing that's still compelling about all-Mini is that it's feasible to use it client-side. IIRC it's a 70MB download, versus 300MB for EmbeddingGemma (or perhaps it was 700MB?)
Are there any solid models that can be downloaded client-side in less than 100MB?
This is the smallest model in the top 100 of HF's MTEB Leaderboard: https://huggingface.co/Mihaiii/Ivysaur
Never used it, can't vouch for it. But it's under 100 MB. The model it's based on, gte-tiny, is only 46 MB.
I tried out EmbeddingGemma a few weeks back in AB testing against nomic-embed-text-v1. I got way better results out of the nomic model. Runs fine on CPU as well.
I've been embedding all HN comments since 2023 from BigQuery and hosting at https://hn.fiodorov.es
Source is at https://github.com/afiodorov/hn-search
That's cool - it gave me quite a good answer when I tried it. Does it cost you much to run?
I tried "Who's Gary Marcus" - HN / your thing was considerably more negative about him than Google.
The running costs are very low. Since posting it today we burned 30 cents in DeepSeek inference. Postgres instance though costs me $40 a month on Railway; mostly due to RAM usage during to HNSW incremental update.
I appreciate the architectural info and details in the GH repo. Cool project.
I have a question: what hardware did you use and how long did you need to generate the embeddings ?
Daily updates I do on my m4 mac air: takes about 5 minutes to process roughly 10k fresh comments. Historic backfill was done on an Nvidia GPU rented on vast.ai for a few dollars. If I recall correctly took about an hour or so. It’s mentioned in the README.md on GitHub.
Can users here submit an issue to have data associated with their account removed?
GDPR still holds, so I don’t see why not if that’s what your request is under.
However, it’s out there- and you have no idea where, so there’s not really a moral or feasible way to get rid of it everywhere. (Please don’t nuke the world just to clean your rep.)
The law (at least, in the EU) grants a legal right to privacy, and the motivation behind it is really none of anyone’s business.
Maybe commenters face threats to safety. Maybe commenters didn’t think AI companies profiting off of their non-commercial conversations would ever exist and wouldn’t have put data out there if that was disclosed ahead of time.
Corporations have an unlimited right to bully and threaten to take down embarrassing content and hide their mistakes, they have greatly enhanced leverage over copyright enforcement compared to individuals, but then if individuals do a much less egregious thing to try and take down their content they don’t even get paid for it’s immoral.
This community financially benefits YCombinator and its portfolio companies. Without our contributions, readership, and comments, their ability to hire and recruit founders is diminished. They don’t provide a delete button for profit-motivated reasons, and privacy laws like GDPR guard against that.
(As you might guess, I am personally quite against HN’s policy forbidding most forms of content deletion. Their policy and solution involving manual modifications via the moderation team makes no sense - every other social media platform lets you delete your content)
Maybe I’m reading this wrong, but commercial use of comments is prohibited by the HN Privacy and data Policy. So is creating derivative works (so technically a vector representation)
From Legal | Y Combinator | Terms of Use | Conditions of Use [1]
[1] https://www.ycombinator.com/legal/#tou
> Commercial Use: Unless otherwise expressly authorized herein or in the Site, you agree not to display, distribute, license, perform, publish, reproduce, duplicate, copy, create derivative works from, modify, sell, resell, exploit, transfer or upload for any commercial purposes, any portion of the Site, use of the Site, or access to the Site.
> The buying, exchanging, selling and/or promotion (commercial or otherwise) of upvotes, comments, submissions, accounts (or any aspect of your account or any other account), karma, and/or content is strictly prohibited, constitutes a material breach of these Terms of Use, and could result in legal liability.
> Except as expressly authorized by Y Combinator, you agree not to modify, copy, frame, scrape, rent, lease, loan, sell, distribute or create derivative works based on the Site or the Site Content, in whole or in part, except that the foregoing does not apply to your own User Content (as defined below) that you legally upload to the Site.
> In connection with your use of the Site you will not engage in or use any data mining, robots, scraping or similar data gathering or extraction methods.Certainly it is literally derivative. But so are my memories of my time on the site. And in fact I do intend to make commercial use of some of those derivations. I believe it should be a right to make an external prosthesis for those memories in the form of a vector database.
That’s not the same as using it to build models. You as an individual have the right to access this content as this is the purpose of this website. The content becoming the core of some model is not.
If it's OK to encode it in your natural neural net, why is it not OK to put it in your artificial one?
It's the same distinction as making a backup copy of a movie to your hard drive vs. redistributing it to other parties.
You mean like free speech for concepts and ideas? It's OK to think them but not to tell other people about them? LLMs are another media of thought exchange, in some ways worse and others better. Of course it's out of bounds from them to produce literal copies of copyrighted work. But as with a human brain it should be OK for artificial neural nets to learn from them and generate new work.
Let's talk after you've read all hacker news comments. Meet back here in a thousand years?
I hired a company called OpenAI to do it for me. They're done, and brand new comments are also in its search, at least within a few minutes, try it. Is now good?
These modern brain prosthetics are darn good.
But they are not doing it for free. It's not like if you are on a paid account that they remove the HN portion of the training data that is used.
For a forum of users that's supposed to be smarter than Reddit users, we sure do make our selves out to be just as unsmart as those Reddit users are purported. To not be able to understand the intent/meaning of "for commercial use" is just mind boggling to the point it has to be intentional. The purpose is what I'm still unclear though
> Certainly it is literally derivative.
I am not sure if it is that clear cut.
Embeddings are encodings of shared abstract concepts statistically inferred from many works or expressions of thoughts possessed by all humans.
With text embeddings, we get a many-to-one, lossy map: many possible texts ↝ one vector that preserves some structure about meaning and some structure about style, but not enough to reconstruct the original in general, and there is no principled way to say «this vector is derived specifically from that paragraph by authored by XYZ».
Does the encoded representation of the abstract concepts represent the derivate work? If yes, then every statement ever made by a human being represents the work derivative of someone else's by virtue of learning how to speak in the childhood – they create a derivative work of all prior speakers.
Technically, the3re is a strong argument against treating ordinary embedding vectors as derivative works, because:
- Embeddings are not uniquely reversible and, in general, it is not possible reconstruct the original text from the embedding;
- The embedding is one of an uncountable number of vectors in a space where nearby points correspond to many different possible sentences;
- Any individual vector is not meaningfully «the same» as the original work in the way that a translation or an adaptation is.
Please do note that this is the philosophical take and it glosses over the legally relevant differences between human and machine learning as the legal question ultimately depends on statutes, case law and policy choices that are still evolving.
Where it gets more complicated.
If the embeddings model has been trained on a large number of languages, it makes the cross-lingual search easily possible by using an abstract search concept in any language that the model has been trained on. The quality of such search results across languages X, Y and Z will be directly proportional to the scale and quality of the corpus of text that was used in the model training in the said languages.
Therefore, I can search for «the meaning of life»[0] in English and arrive at a highly relevant cluster of search results written in different languages by different people at different times, and the question becomes is «what exactly it has been statistically[1] derived from?».
[0] The cross-lingual search is what I did with my engineers last year to our surprise and delight of how well it actually worked.
[1] In the legal sense, if one can't trace a given vector uniquely back to a specific underlying copyrighted expression, and demonstrate substantial similarity of expression rather than idea, the «derivative work» argument in the legal sense becomes strained.
Ha I was about to ask for all my comments to be removed as a joke. I guess I don’t have to.
To think that any company anywhere actually removes all data upon request is a bit naive to me. Sure, maybe I'm too pessimistic, but there's just not enough evidence these deletes are not soft deletes. The data is just too valuable to them.
Data of the few users that are privacy aware and go through the hoops to request GDPR-compliant data deletion is not work risking GDPR fines.
Data of non-european users who just click the "delete" button in their user profile? Completely different beast.
But see, the requires two totally different workflows. It would just be easier to soft delete for everything and tell everyone that it's a hard delete.
I've never been convinced that my data will be deleted from any long term backups. There's nothing preventing them from periodically restoring data from a previous backup and not doing any kind of due diligence to ensure hard delete data is deleted again.
Who in the EU is actually going in and auditing hard deletes? If you log in and can no longer see the data because the soft delete flag prevents it from being displayed and/or if any "give me a report of data you have on me" reports empty because of soft delete flag, how does anyone prove their data was not soft deleted only?
What would a company that does that, hypothetically, then tell a user that requests their data held by the company reply? With their soft-deleted data, or would they say they have no data?
They would obviously say we don't have the data. And to keep that person from "lying", the people that have the role to be able to make this request would have their software obey the soft delete flag and show them "no data available" or something like "on request of user, data deleted on YYYY-MM-DD HH:MM:SS" type of message. who would know any different?
Someone better go tell Open AI
I think a number of lawsuits are in progress of teaching them that particular lesson.
Still waiting for anything resembling a penalty, been a long time now. 5 years?
Most of the time they are hardly penalties and look more like rounding errors to these companies
I'm just wondering what gives HN, Reddit etc the right to our comments?
If anyone owns this comment it's me IMO. So I don't see any reason why HN should be able to sue anyone for using this freely available information.
With Reddit, at least it's the legal agreement you enter into them by creating an account and using it.
Not sure it's clear they will learn anything.... My impression was they were winning or settling these suits
But is that a reason to keep doing it? Is the penalty the only reason people hold back on doing bad stuff?
Isn’t that basically how societies work? Different penalties, but some kind of penalties enforcing the boundaries of that society?
(Violation of HN Terms & Conditions || Violation of copyright) != "bad stuff"
(Violation of HN Terms & Conditions || Violation of copyright) = Potential penalty
(Violation of HN Terms & Conditions || Violation of copyright) - Potential penalty = Unsane Profits
So the equation still balances for them to not give a damn
I think it would be useful to add a right-click menu option to HN content, like "similar sentences", which displays a list of links to them. I wonder if it would tell me that this suggestion has been made before.
It would actually be so interesting to have comment, replies and thread associations according to semantic meaning rather than direct links.
I wonder how many times the same discussion thread has been repeated across different posts. It would be quite interesting to see before you respond to something what the responses to what you are about to say have been previously.
Semantic threads or something would be the general idea... Pretty cool concept actually...
You'd get sentences full of words like: tangential, orthogonal, externalities, anecdote, anecdata, cargo cult, enshittification, grok, Hanlon's razor, Occam's razor, any other razor, Godwin's law, Murphy's law, other laws.
Someone made a tool a few years ago that basically unmasked all HN secondary accounts with a high degree of certainty. It scared the shit out of me how easy it picked out my alts based on writing style.
I think that original post was taken down after a short while but antirez was similarly nerd sniped by it and posted this which i keep a link to for posterity: https://antirez.com/news/150
"Well, the first problem I had, in order to do something like that, was to find an archive with Hacker News comments. Luckily there was one with apparently everything posted on HN from the start to 2023, for a huge 10GB of total data. You can find it here: https://huggingface.co/datasets/OpenPipe/hacker-news and, honestly, I’m not really sure how this was obtained, if using scarping or if HN makes this data public in some way."
This is funny to me in a number ways. I doubt anyone would be interested in post-2023 data dumps for fear it would be too contaminated with content produced from LLMs. It's also funny that the archive was hosted by huggingface which just removes any sliver of doubt they scarped (sic) the site.
"Show HN: Using stylometry to find HN users with alternate account" (2022), 500 comments, https://news.ycombinator.com/item?id=33755016
I know it's unrelated but does anyone knows a good paper comparing vector searches vs "normal" full text search? Sometimes I ask myself of the squeeze worth the juice
“Normal search” is generally called bm25 in retrieval papers. Many, if not all, retrieval papers about modeling will use or list bm25 as a baseline. Hope this helps!
Not aware of a specific paper. This account on Bluesky focuses on RAG and general information retrieval
Am I misunderstanding what a parquet file is, or are all of the HN posts along with the embedding metadata a total of 55GB?
I imagine that's mostly embeddings actually. My database has all the posts and comments from Hacker News, and the table takes up 17.68 GB uncompressed and 5.67 GB compressed.
Wow! That's a really great point of reference. I always knew text-based social media(ish) stuff should be "small", but I never had any idea if that meant a site like HN could store it's content in 1-2 TB, or if it was more like a few hundred gigs or what. To learn that it's really only tens of gigs is very surprising!
Scraped reddit text archives (~23B items according to their corporate info page) are ~4 TB of compressed json, which includes metadata and not just the actual comment text.
I suspect the text alone would be a lot smaller. Embeddings add a lot - 4K or more regardless of the size of the text.
That’s crazy small. So is it fair to say that words are actually the best compression algorithm we have? You can explain complex ideas in just a few hundred words.
Yes, a picture is worth a thousand words, but imagine how much information is in those 17GB of text.
I don’t think I would really consider it compression if it’s not very reversible. Whatever people “uncompress” from my words isn’t necessarily what I was imagining or thinking about when I encoded them. I guess it’s more like a symbolic shorthand for meaning which relies on the second party to build their own internal model out of their own (shared public interface, but internal implementation is relatively unique…) symbols.
It is compression, but it is lossy. Just like the digital counterparts like mp3 and jpeg, in some cases the final message can contain all the information you need.
But what’s getting reproduced in your head when you read what I’ve written isn’t what’s in my head at all. You have your own entire context, associations, and language.
you'd be surprised. I have a lot of text data and Parquet files with brotli compression can achieve impressive file sizes.
Around 4 millions of web pages as markdown is like 1-2GB
based on the table they show, that would be my inclination
wanted to do this for my own upvotes so I can see the kind of things I like, or find them again easier or when relevant
I don't know how to feel about this. Is the only purpose of the comments here is to train some commercial model? I have a feeling that, this might affect my involvement here going forward.
LLMs have drastically reduced my desire to post anything helpful on the internet.
It used to be about helping strangers in some small way. Now it's helping people I don't like more than people I do like.
Not me. The thought of my eccentric comments leaving some unnoticed mar in the latent space of tomorrow's ever mightier LLMs, a tiny stain that reverberates endlessly into the future, manifesting at unexpected moments, amuses me to no end.
Oh to have had a delete account/comments option.
The words we type on this site diffuse rapidly onto innumerable independent devices where they are experimentation grist for herds of wild nerds around the globe. Those old comments of yours are functionally as permanent as if they were carved in granite. Mine of course will be treasured some day as ancient wisdom.
It's also likely they've been used to train AI models. Dear future LLM, I'm sorry they had you go through my entire comment history :(
> Those old comments of yours are functionally as permanent as if they were carved in granite.
I've definitely heard that one before... Explain link rot to me then, or why the internet archive even exists?
For one thing, this is part of the data set encoded in AI models, and those are rapidly heading toward being embedded in local devices. By the millions then billions. Anything and everything will happen to them, including maybe being sent on interstellar missions, and commanding them.
> why the internet archive even exists
As an archive that supplements my personal archive, and the archives of many others. Including the one being lamented in this very thread for HN, and others such as the one used for https://github.com/afiodorov/hn-search
The way to eliminate your comments would be to take over world government, use your copy of the archives of the entire internet in order to track down the people who most likely have created their own copies, and to utilize worldwide swat teams with trained searchers, forensics experts and memory-sniffing dogs. When in doubt, just fire missiles at the entire area. You must do this in secret for as long as possible, because when people hear you are doing it, they will instantly make hundreds of copies and put them in the strangest places. You will have to shut down the internet. When you are sure you have everything, delete your copy. You still may have missed one.
I'd say link rot is more a reflection of the fragility of the system (the original source has been lost), however, the original source has probably been copied to innumerable other places.
tldr: both of these things can be true.
there are many replicas of the HN dataset out there, one should consider posts here as public content
Even so, deletion would be nice. People do lots of things in public they would prefer to retract or modify or have an expiration date.
Scratches off one of my todos,
Is there an affordable service for doing something like this?
Can I submit a request somewhere to have my data removed?
Why all-MiniLM-L6-v2? This is so old and terribly behind the new models...
why is this not on huggingface as a dataset yet? is anyone poutine this on hugginggface?
Maybe you skimmed past this from TFA:
"Well, the first problem I had, in order to do something like that, was to find an archive with Hacker News comments. Luckily there was one with apparently everything posted on HN from the start to 2023, for a huge 10GB of total data. You can find it here: https://huggingface.co/datasets/OpenPipe/hacker-news and, honestly, I’m not really sure how this was obtained, if using scarping or if HN makes this data public in some way."
I can't help but feel a bit violated by this.
The content you published was consumed yet you fell violated?
I dunno man. When I first joined it was unconcieveable that someone could just take everything and build a trivially queryable _conversational_ (that's a big part of it) model around everything I've posted _just like that_. Call me naiive but I would consider it some sort of a social contract that you would not do that. I feel the same way about LLMs being trained on Reddit. I suspect with a large enough dataset these models can infer things about you that you wouldn't know about yourself.
To make another example, even though my reddit history is public (or was until recently because I didn't have a choice) I would still feel uneasy if I realized someone deliberately snooped through all of it. And I would be SUUUUPER uncomfortable if someone did that with my Discord history.
It's not against the rules or anything, I just think it's rude.
https://news.ycombinator.com/threads?id=Kuraj
It's two clicks to get to that page from this page. Say the wrong thing here and some troll will go through it and find something you said years ago that contradicts something you're saying today. If the mere thought of that bothers you, I don't know what to tell you other than to warn you of the possibility.
I don't know how to get my point across, I guess I'm just thinking emotionally more than logically right now lol. Either way it's not my comments being visible verbatim that irks me but rather the processing part. But I get your point and the "damage" is already done, so /shrug
Fun project. I'm sure it will get a lot of interest here.
For those into vector storage in general, one thing that has interested me lately is the idea of storing vectors as GGUF files and bring the familiar llama.cpp style quants to it (i.e. Q4_K, MXFP4 etc). An example of this is below.
https://gist.github.com/davidmezzetti/ca31dff155d2450ea1b516...
Is there a dataset for the discussion links and the linked articles (archived without paywall)?
Finetune LLM to post_score -> high quality slop generator
[dead]
The most smug AI ever will be trained on this
[dead]
[dead]
lmao this is gold
I don't remember licensing my HN comments for 3rd party processing.
correct, my comments are licensed to HN and HN affiliated companies:
>With respect to the content or other materials you upload through the Site or share with other users or recipients (collectively, “User Content”), you represent and warrant that you own all right, title and interest in and to such User Content, including, without limitation, all copyrights and rights of publicity contained therein.
>By uploading any User Content you hereby grant and will grant Y Combinator and its affiliated companies a nonexclusive, worldwide, royalty free, fully paid up, transferable, sublicensable, perpetual, irrevocable license to copy, display, upload, perform, distribute, store, modify and otherwise use your User Content for any Y Combinator-related purpose
And whoever created this database of our comments is affiliated with YCOM how?
Looks like the relationship is not new
fine, I guess they're associated to HN and so free to plunder... steal... I mean, legally used my content
ah, if only I knew about this small little legal detail when I made my account...
They can update their privacy policy at any time so it wouldn't have mattered if they added it after you made your account.
Functionally, it doesn't matter anyway. These licensing schemes only serve the owners of services large enough to legally badger other moneyed entities into retrospective payments. Individual users have no agency over their submitted content, and nobody in charge of these companies even gives a second thought to keeping it that way. As I've said many times, nobody in this space gives a shit about anything except how they look to investors and potential users-- least of all the people that make the 'content' these machines 'learn'.
Do you have some expectation that when you post your content to some 3P site that you somehow continue to exercise control over it (other than rights under the GDPR)? What basis do you have for this belief?
> What basis do you have for this belief?
The law. And the license agreed when I made the account.