gaoj0017 commented on PR #14078:
URL: https://github.com/apache/lucene/pull/14078#issuecomment-2644556568

   After Elastic’s last round of replies, the Elastic team reached us for 
clarification on the issues via zoom meetings. In the meetings, they promised 
to fix the misattribution, so we suspended our reply and waited for their 
amendment. Recently, Elastic made some edits on the [BBQ 
blog](https://www.elastic.co/search-labs/blog/better-binary-quantization-lucene-elasticsearch)
 and this pull request. However, (1) the updated blog still positions BBQ as a 
new method (despite that BBQ majorly follows our RaBitQ method with some 
differences in the implementation only), and (2) the updated PR does not 
position our extended RaBitQ method as a prior art of their method (despite 
that the method is similar to our extended RaBitQ method at its core and our 
method is a prior art). As academic researchers, we feel tired of writing 
emails asking for proper attribution round after round, and hence we would like 
to speak out for ourselves through this channel here.
   
   Point 1 in our reply: This pull request is a direct follow-up to [the 
preceding pull request](https://github.com/apache/lucene/pull/13651) of the 
so-called “BBQ” method. Thus, this point should not be ignored and excluded 
from the discussion. We emphasize again that all the key features of the 
so-called “BBQ” method follow our RaBitQ method, with only minor modifications. 
 
   
   Point 2 in our last reply: For the community’s information, we would like to 
provide the following timeline. 
   
   - On 16-Sep-2024, we published our [extended RaBitQ 
paper](https://arxiv.org/abs/2409.09913) along with codes. 
   - On 25-Oct-2024, we circulated the extended RaBitQ paper to Elastic as we 
knew they were implementing RaBitQ in Lucene. They replied positively that they 
will eagerly read our new paper. At that time, we didn’t expect that they would 
announce a variant of our RaBitQ method as their own innovation later. In 
addition, it is worth highlighting that our extended RaBitQ method is the first 
to achieve reasonable accuracy with 1-bit and 2-bit binary/scalar quantization 
in the literature by then.
   - On 11-Nov-2024, Elastic announced the BBQ method as their new innovation 
(despite that BBQ is simply our RaBitQ method with some differences in 
implementation).
   - On 22-Nov-2024, we expressed our concerns over the BBQ method to Elastic 
via email. Elastic replied that they would need internal discussions and would 
get in touch soon. 
   - On 14-Dec-2024, we had not received any replies from Elastic and Elastic 
kept marketing the so-called “BBQ” method as their own innovation in several 
blogs and press releases. We therefore expressed our concerns over the BBQ 
method to the Lucene community (https://github.com/apache/lucene/pull/13651).
   - On 19/20-Dec-2024, Elastic announced the OSQ method in this PR and [this 
blog](https://www.elastic.co/search-labs/blog/scalar-quantization-optimization).
 They claimed to have unlocked the performance gain for various bits. However, 
they ignored the fact that our extended RaBitQ method can already handle 
various bits, despite that the method was shared with Elastic on 25 Oct 2024 
and published on arXiv in Sep 2024. 
   - 14-Dec-2024 to 07-Jan-2025, Elastic and we had a few rounds of 
communications via this channel at GitHub, which are visible to the public. The 
last response was from Elastic on 07-Jan-2025, and we stopped replying here 
because Elastic reached out to us after their response and proposed to have 
communications via other channels (email and meetings) and we agreed with the 
hope that Elastic would fix the acknowledgement problem.
   - 13-Jan-2025 and 22-Jan-2025, Elastic and we had two Zoom meetings and 
email communications and discussed how to provide sufficient acknowledgements 
to RaBitQ and its extension. Afterwards, Elastic made some edits on [the BBQ 
blog](https://www.elastic.co/search-labs/blog/better-binary-quantization-lucene-elasticsearch)
 and this PR. However, (1) in the updated blog, Elastic still positions BBQ as 
a new methodology and does not explicitly specify that the core ideas of BBQ 
follow those of RaBitQ and (2) in this PR, Elastic still positions OSQ as a 
parallel work of our extended RaBitQ method (despite its core ideas are similar 
to our extended RaBitQ method and our method is a prior part and published 
months ago). In addition, Elastic shared a doc with us explaining the 
differences between OSQ and extended RaBitQ, but our understanding is that the 
core ideas of two methods, i.e.,  trying different parameters for scalar 
quantization on a per-vector basis, are similar.
   - 05-Feb-2025, this is when Elastic contacted us via <ins>Email</ins> lastly 
(as of today), saying that they think the acknowledgement of our RaBitQ and 
extended RaBitQ (based on the updated blog and PR) is sufficient. 
   - 08-Feb-2025 (today), this is when we replied to Elastic via 
<ins>Email</ins> lastly, updating them we shall speak out for us in other 
channels (after getting tired of writing emails round after round). 
   
   In summary, our requirements are two-fold: (1) Elastic should position BBQ 
as a variant of RaBitQ but not as a new innovation; and (2) Elastic should 
acknowledge our extended RaBitQ method as a prior art, which has covered 
similar ideas of the OSQ method and been published months earlier, but not a 
parallel work. 
   
   As for future communications, we may not be able to respond to every future 
response from Elastic because we feel they can keep arguing and changing their 
arguing points - in this thread, Elastic have changed their attribution of OSQ 
several times (from the inspiration of RaBitQ to LVQ and SCANN, and to their 
own earlier PR). We should have been focusing on research more as academic 
researchers. No matter how they would argue, they cannot deny the following two 
points: 
   1. <ins>**The so-called BBQ majorly follows our RaBitQ method;** </ins> 
   2. <ins>**The OSQ method (introduced in this PR) has its major idea similar 
to our extended RaBitQ method and our extended RaBitQ method is a prior art 
which achieves good accuracy at 1-bit/2-bit binary/scalar quantization for the 
first time.**</ins> 
   
   We would also consider reaching out to more channels to speak out for 
ourselves if necessary.
   
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org

For queries about this service, please contact Infrastructure at:
us...@infra.apache.org


---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org
For additional commands, e-mail: issues-h...@lucene.apache.org

Reply via email to