jkb79

6 ms·

1.
by jkb79 3y ago
ColBER: Contextualized Late Interaction over BERT. That is just the name. It can be fine-tuned for retrieval using any encoder-only model like the ones you mention.
2.
Advent of Tensors 2023 – 24 challenges(blog.vespa.ai)
3 points by jkb79 3y ago|1 comments
3.
by jkb79 3y ago
Yes, Vespa has a generic Tensor framework that allows you to index multiple vectors for a single field, see https://blog.vespa.ai/semantic-search-with-multi-vector-inde... for details. field embeddings type tensor<float&
4.
by jkb79 3y ago
Actually, Vespa comes out of the same FAST company. Yahoo bought Overture/Altavista and a lot of other web search companies in 2003, including the web search division of FAST. The Enterprise search division of FAST was later acquired b
5.
by jkb79 3y ago
Hehe, it was a joke, we don't have polar bears on the mainland of Norway. But, it was fun to show the photo to visitors from different countries. "Be careful when you walk back to the hotel".
6.
by jkb79 3y ago
Disclaimer, I work on Vespa. If you look for just pure vector similarity search, there are many alternatives. But Vespa's tensor support, multi-vector indexing and the ability to express models like colBERT ( 1) or cross-encoders makes
7.
by jkb79 3y ago
Vespa was the internal code name for the project going back to 2005ish, vertical search platform.
8.
by jkb79 3y ago
Thank you for the shout-out Frank!
9.
by jkb79 3y ago
Yeah, I worked with the Flickr team on that project. Scaling to billions of photos, with partial update support of popularity for ranking. Back then, the properties had to stand up their own Vespa cluster(s), later on we created a managed s
10.
by jkb79 3y ago
>I don't see a ton of data that points towards us (vector DB ppl) building >towards traditional TREC/BEIR #s. This is highly accurate, most vector database companies don't talk about the shortcomings of vector represent
11.
by jkb79 3y ago
It's an opinionated blog post published on Arxiv, masquerading as research. IMHO, it's a gigantic self-own and doesn’t promote Lucene in a good way. For example, by demonstrating how they get only 10 QPS out of a system with 1TB o
12.
by jkb79 3y ago
Very good point, plus that if you introduce approximate search instead of exact search, you can only loose quality and in some cases the loss is so significant that it becomes worse than plain simple techniques.
13.
Accelerating Transformer-Based Embedding Retrieval with Vespa(blog.vespa.ai)
2 points by jkb79 3y ago|0 comments
14.
by jkb79 3y ago
Faiss implements multiple different ANN algorithms, and some like IVF can use disk based indexes.
15.
by jkb79 3y ago
Disclaimer, I'm a developer working on the Vespa.ai project. One reason is that we simply don't have DevRel teams or marketing teams, but we still have decent interest, from large companies like Spotify using Vespa in production f
16.
by jkb79 3y ago
That is a very good analogy!
17.
by jkb79 3y ago
Only Vespa allows you to index multiple vectors per schema field, avoiding duplicating all the meta data of the document into the "chunk", and avoids maintaining the document to chunk fan-out. See https://blog.vespa.ai
18.
Customizing Reusable Frozen ML-Embeddings with Vespa(blog.vespa.ai)
1 points by jkb79 3y ago|0 comments
19.
Building a billion-scale vector search solution without breaking the bank(blog.vespa.ai)
2 points by jkb79 4y ago|0 comments
20.
by jkb79 4y ago
Yes, https://blog.vespa.ai/vespa-hybrid-billion-scale-vector-sear...
21.
by jkb79 4y ago
Friendly reminder, code that you can read is not open source. The particular repo does not have any license information and cannot be considered open source.
22.
Vector Podcast – Journey of Vespa from Sparse into Neural Search(youtube.com)
3 points by jkb79 4y ago|1 comments
23.
by jkb79 5y ago
Interesting to see how deep learning is applied to search using learned vector representations of documents (podcasts) and user queries.
24.
Spotify Introducing Natural Language Search(engineering.atspotify.com)
6 points by jkb79 5y ago|1 comments
25.
by jkb79 5y ago
Great blog post, throughput is very often overlooked when talking about model inference.
26.
by jkb79 5y ago
>Why does 100M vectors not work in ES? I think that it's important to mention Elasticsearch version in this context :=) Full disclosure, I work on Vespa.ai, but I have a pretty good understanding of how NN and ANN works in ES. Elast
27.
Show HN: Bert NLP inference in browser using WebAssembly-SIMD(github.com)
2 points by jkb79 5y ago|0 comments
28.
by jkb79 5y ago
MS Marco is the largest relevancy collection in the open, see https://microsoft.github.io/msmarco/
29.
by jkb79 5y ago
Vector models are nothing but representation learning and applying the model out-of-domain usually gives worse results than plain old BM25. See https://arxiv.org/abs/2104.08663 A concrete example is DPR which is a stat
30.
by jkb79 5y ago
>A much more interesting (and harder) problem is creating good vectors to begin >with. Indeed, this is the hardest problem. Vector search shines when used in-domain using deep representation learning, for example bi-encoders on top of

More ›