```
We evaluated several precision pairings across our internal retrieval benchmark suite. Scores are NDCG@10 averaged across the suite, scaled to 0–100. NDCG@10 (Normalized Discounted Cumulative Gain at rank 10) measures how well the top 10 results are ordered against the ideal ranking, rewarding relevant documents more when they appear higher, with 100 being a perfect ranking. The full-precision baseline averages 90.26. Int8 query against binary documents averages 89.65, a 0.61 point drop, while reducing document-vector storage by 32x
```
Saying "Near lossless" to mean 90% accurate retrieval of saved vectors is simply a lie. Lossy-ness is binary, not something you can paper over with getting close enough. And 90% is not close. Sure, LLMs are all about gradient descent on noisy data sets so I guess this is acceptable in this field but that terminology usage still bothered me
I don't believe that's what they were saying at all though. The claim appears to be that it's near lossless relative to their own baseline that uses float. Which I'd grant, since a 32x storage reduction for 0.61% loss in quality is a reasonable trade off when you've already decided to accept that ~90% is "good enough".
I would love to see real examples of what reduced quality means in practice. Are you able to recover a document from the vector in a human readable format? If so, what sort of changes come up?
I could imagine a scenario where differences tend to be more substantive than you'd expect because of how less frequent words with fine distinctions in meaning - the very words that make the document special - may be embedded in the vector space.
Most of the fine distinctions are already lost when a document is processed through a pile of linear algebra to turn it into a fixed-size list of floating-point numbers, as you can see from the NDCG@10. Vector search is not a tool for fine distinctions. It's a tool for reducing a large pile of documents to a smaller selection of candidates, which you can then check individually with some more expensive method.
this is the reason why we report ndcg and not recall. ndcg respects fine grained details so you get the an overview of how much details you are trading off since it would hurt the ranking.
Asymmetry is clever. FWIW, this is very similar to the strategy employed by BitNet models (i.e., int8 activations with binary or ternary weights); I suspect retrieval is a little more amenable to this approach.
In principle, binary x binary should be pretty fast since it just requires bitwise XNOR and popcount/reduction, but in practice it's slow unless you've really optimized it. And, as stated in the article, you'd still be losing a lot of accuracy that way.
Hey breadislove; amazing article, I’ll be sending mixedbread an email in the morning that may interest you (email will be <5-characters>@pm.me)
I have also been working in compression and performance engineering, and managed to get a 99+% compression unlock versus conventional approaches (100+KB down to 1KB) in the scenario of 30 minute massive multiplayer game replays for a “game+engine” I’m developing
I think there’s a synergy between these 2 concepts I’d love to chat some more
There is, after you define what you’re ready to loose and understand the lossy space. That’s how we came up with mobile cellphones, audio and video codecs etc. Literally powering all modern devices we use.
``` We evaluated several precision pairings across our internal retrieval benchmark suite. Scores are NDCG@10 averaged across the suite, scaled to 0–100. NDCG@10 (Normalized Discounted Cumulative Gain at rank 10) measures how well the top 10 results are ordered against the ideal ranking, rewarding relevant documents more when they appear higher, with 100 being a perfect ranking. The full-precision baseline averages 90.26. Int8 query against binary documents averages 89.65, a 0.61 point drop, while reducing document-vector storage by 32x ```
Saying "Near lossless" to mean 90% accurate retrieval of saved vectors is simply a lie. Lossy-ness is binary, not something you can paper over with getting close enough. And 90% is not close. Sure, LLMs are all about gradient descent on noisy data sets so I guess this is acceptable in this field but that terminology usage still bothered me
I don't believe that's what they were saying at all though. The claim appears to be that it's near lossless relative to their own baseline that uses float. Which I'd grant, since a 32x storage reduction for 0.61% loss in quality is a reasonable trade off when you've already decided to accept that ~90% is "good enough".
near lossless refers to being 89.65/90.26 = 99.32% of baseline, i'm pretty sure.
yes exactly.
I would love to see real examples of what reduced quality means in practice. Are you able to recover a document from the vector in a human readable format? If so, what sort of changes come up?
I could imagine a scenario where differences tend to be more substantive than you'd expect because of how less frequent words with fine distinctions in meaning - the very words that make the document special - may be embedded in the vector space.
Most of the fine distinctions are already lost when a document is processed through a pile of linear algebra to turn it into a fixed-size list of floating-point numbers, as you can see from the NDCG@10. Vector search is not a tool for fine distinctions. It's a tool for reducing a large pile of documents to a smaller selection of candidates, which you can then check individually with some more expensive method.
The ndcg loss is minimal 90.26 -> 89.65. This means it maintains most of the quality.
this is the reason why we report ndcg and not recall. ndcg respects fine grained details so you get the an overview of how much details you are trading off since it would hurt the ranking.
Asymmetry is clever. FWIW, this is very similar to the strategy employed by BitNet models (i.e., int8 activations with binary or ternary weights); I suspect retrieval is a little more amenable to this approach.
In principle, binary x binary should be pretty fast since it just requires bitwise XNOR and popcount/reduction, but in practice it's slow unless you've really optimized it. And, as stated in the article, you'd still be losing a lot of accuracy that way.
Hey breadislove; amazing article, I’ll be sending mixedbread an email in the morning that may interest you (email will be <5-characters>@pm.me)
I have also been working in compression and performance engineering, and managed to get a 99+% compression unlock versus conventional approaches (100+KB down to 1KB) in the scenario of 30 minute massive multiplayer game replays for a “game+engine” I’m developing
I think there’s a synergy between these 2 concepts I’d love to chat some more
Any way I can read about this or the use case? I have a hobby interest
to which email did you send it? can u send it to support please?
To people smarter than me, how impressive and/or revolutionary is this?
there is no such thing as "near lossless"
There is, after you define what you’re ready to loose and understand the lossy space. That’s how we came up with mobile cellphones, audio and video codecs etc. Literally powering all modern devices we use.
So then ... "lossy"
3 replies →
Actually, all of those things are considered "lossy".
1 reply →
If you squint hard enough, it sounds like their storage layer is a bloom filter
The Pi compression algorithm is better.
Doubtful. The problem with the pi idea is that you need to include the offset, which will likely be as long as or longer than your data.
" A single document produces more then one embedding, depending on the complexity of the document it can produce hundreds or thousands of vectors."
That typo up there is kind of endearing in the AI slop era.
Not seeing a typo in your quote. Can you point it out?
I think they're referring to "then" vs "than"
1 reply →
[flagged]
[dead]
[dead]
[dead]
[flagged]
[flagged]
[flagged]
Unfortunately as cost reduction trends to 100%, it comes along with an intrinsic high-pass sarcasm filter.
You would obviously be trading storage for compute and time to retrieve the storage.
100% reduction is impossible for something which should work, because -100% means it is now 0
They were clearly being sarcastic
Reminds me of 'Learning to be me' by Greg Egan