vespa.ai
@vespaengine
Followers
3K
Following
772
Media
29
Statuses
514
https://t.co/abkb8IjPSH - the open source platform for combining data and AI, online. Vectors/tensors, full-text, structured data; ML model inference at scale.
Joined September 2017
Announcing: The RAG Blueprint. Build RAG like the world's most successful applications. Start from our open source sample app which contains all you need to do to achieve world-class quality at any scale. Sample app: Blog post:
1
3
24
New Vespa features covered in the June newsletter:.- Layered ranking: Rank chunks in documents. - Elementwise bm25.- top, filter_subspaces, and cell_order tensor functions.- chunking support in indexing.- element-gap: Proximity over chunks.- filtering in grouping results.-.
1
4
11
With layered ranking, you can use Vespa's powerful ranking machinery both to select documents and then to select the right chunks within them.
1
0
7
Until now you've either had to .- index document chunks as separate documents, creating a billion documents with no context, or.- index entire documents with many chunks, preserving context but feeding too much noise to LLMs.
1
1
7
RT @radu0gheorghe: If you want to learn more about @vespaengine, you might find our playlists interesting. Lots of podcasts and conference….
0
2
0
Recommended weekend listening is this podcast from AWS with our CEO.
My podcast with AWS is out Some of what we talked about:. - Even a superintelligent LLM won't help if you can't give it the right data. - This is called "relevance", and is Not Exactly a New Problem. - With deep research we're seeing query load exploding.
0
0
2
We've updated our ES comparison to cover Elastic 9. Congrats to Elastic, achieving latencies just 3x those of Vespa is no small feat!.
TL;DR. 1. #Elasticsearch 9 is more efficient than 8, gap to @vespaengine reduced to ~3x with 16 clients.2. Single client latency is higher, unless force-merged => a bigger gap (~1.7x for hybrid).3. Pushing more load increases both gaps:.- ES 9 >> ES 8.- Vespa >> ES 9.
0
0
10
RT @andreer: cool to learn @allen_ai are using @vespaengine! and binarized embeddings are amazing for cost/perf, this is a great model.
0
3
0
RT @ravo: Spreading Vespa cheer in Prague today — wearing what we run. powers the @searchplex stack and my holida….
0
2
0
We'll have time to explain later, but for those who understand, enjoy Vespa 8.530.
3
4
22
Another, related new feature: Proximity between chunks. If you index chunks, setting this will instantly improve your quality.
This is a cool addition. It's more than just chunking too. Generally, lexical search algorithms like BM25 and TF-IDF are tailored for a world of whole documents. Then, lots of modern embeddings and semantic retrieval benefit from smaller text chunks (and maybe prefix/suffix.
0
0
9
As most have moved to multi-chunk documents in RAG applications, we thought we should make this easier. Introducing the chunk indexing function:
2
3
26
RT @jonbratseth: According to the recent Columbia Journalism Review, Perplexity has the best AI Search and only ChatGpt is even in the same….
0
2
0
RT @thomas_thoresen: Hard problem indeed. Not everyone wraps google or start from scratch though. Case in point: @perplexity_ai builds o….
0
2
0
RT @radu0gheorghe: Introduction to @vespaengine (with lots of references) for #Solr users: As always, feedback is….
0
2
0
The actions of an LLM will never be better than the information you give it to work with. Those who understand this, starts to take information retrieval very seriously. What does that look like? A short thread 🧵.
0
0
4