Vespa.ai - the open source platform for combining data and AI, online. Vectors/tensors, full-text, structured data; ML model inference at scale.
- Tweets 591
- Following 3
- Followers 3,601
- Likes 872
The Singaporean government has deployed Vespa to search every word ever said in their Parliament.
"A good decision is an informed one [...] The heart of a good RAG system is a good search engine to retrieve the relevant data chunks for ingestion"
Many teams are racing to make use of these new methods to achieve superior quality, but the Singaporean government may have been first to put them in production!
3
15
84
11,652
We have become our own company!
Expect even more features, even faster.
blog.vespa.ai/vespa-is-becom…
4
14
61
4,200
Announcing the open sourcing of Vespa, Yahoo’s Big Data Processing and Serving Engine at vespa.ai: blog.vespa.ai/post/165763618…
3
61
55
Announcing our $31 million raise from Blossom Capital blog.vespa.ai/announcing-our…
We'll spend it all on features.
1
5
44
4,048
Spotify launches semantic search in Podcasts, powered by vespa.ai
Here’s the problem: you want to search for a podcast, but you can’t remember the name, only what it’s about.😖 Here comes Natural Search 🔍 to save the day!
engineering.atspotify.com/in…
5
42
LangChain Webinar on retrieval for LLMs with @lateinteraction and our very own @jobergum in six hours.
Save your spot at crowdcast.io/c/vs7bhtxbersr
1
5
30
4,662
Choosing an algorithm for fast vector search for big data serving
medium.com/vespa/approximate…
Read if you are interested in combining fast vector search with filters, text, and real-time updates.
Or in how professionals choose algorithms from the literature for production usage.
12
30
Having trouble keeping up?
Guidebook to the State-of-the-Art Embeddings and Information Retrieval
by @aapo_tanskanen at @thoughworks is out today - a great resource to get up to date.
linkedin.com/pulse/guidebook…
8
26
2,446
Search is going through a paradigm shift -
neural methods are outperforming traditional methods by a wide margin.
Can you use it real production systems?
[1/4]
1
8
26
What the world need most now is more research on #covid19, faster. We've created cord19.vespa.ai to help with that.
It lets researchers find research papers by combining text and structured search with exploring by semantic similarity using the scibert-nli model.
The @vespaengine team released cord19.vespa.ai based on the CORD-19 dataset released by the @allen_ai.
Since everything is open-sourced, you can contribute to the project in multiple ways. 👇
#NLP #NLProc #SearchEngine #COVIDー19
medium.com/@thigm/vespa-ai-a…
14
24
Do you want to work on
- search relevance
- recommendation
- personalization
using
- machine-learning
- embeddings
- vector or hybrid retrieval
but are kinda tired of all the plumbing?
Thread time 👇
2
7
24
After extensive benchmarking, Marqo migrated their underlying engine from OpenSearch to Vespa:
"For Marqo 2, we looked at a number of open source and proprietary vector databases, including Milvus, Vespa, OpenSearch (AWS managed), Weaviate, Redis, and Qdrant."
1
23
2,181
We made it simpler to create semantic search applications on Vespa
- No need to create vectors yourself
- No code needed
- No limits, once you want to grow add features
blog.vespa.ai/text-embedding…
3
2
24
Introducing TensorFlow support
blog.vespa.ai/post/171861434…
12
21
We're starting a new series of posts where we provide a complete and scalable Vespa application with a frontend for a popular use case.
First out: Shopping sites!
blog.vespa.ai/post/184617258…
8
19
Follow this guide if you need to build a personal data app which
- costs about 5% of a traditional vector database
- contrary to those, finds all the user's data
- scales to millions (or billions) of users
Hands-on RAG guide for personal data with @vespaengine and @llama_index 😍
Featuring LLamaIndex retrievers, Vespa streaming mode for personal data, built-in embedders, multi-index RAG (federation + blend), hybrid search, and rank fusion strategies.
blog.vespa.ai/scaling-person…
1
3
18
2,175
Releasing support for ONNX models (Caffe2, PyTorch). All the same optimizations making it fast as TensorFlow models since we compile into the same execution engine blog.vespa.ai/post/175233055…
9
18
Announcing support for Maximum Inner Product Search (MIPS) in Vespa.
✅ Search efficiently for vectors with the highest dot product, by an innovative extension to HNSW indexes.
✅ With full realtime indexing, and no need to normalize vectors.
blog.vespa.ai/announcing-max…
3
18
2,257
Announcing approximate nearest neighbor vector search in Vespa blog.vespa.ai/vespa-product-…
- State of the art performance from a native Vespa implementation of HNSW
- Vector search combines efficiently with filters and text search
- Infinitely scalable
- Real-time updates of vectors
12
19
Things we've been up to in May:
👓LLM integration: Build complete RAG applications on Vespa.
🎰Embed to multiple representations at once: An oom cheaper vector search by creating a small vector for search and a larger for ranking.
🎯Combine fuzzy search and prefix match: Ideal for typeahead search.
🚴10x faster export-import between Vespa clusters.
🐎30-900% faster vector search from distance calculation and conversion optimizations.
🐍Lots of new pyVespa features.
1
3
18
5,318
Recommender systems need to multiply very large sparse matrices. e-commerce platform @farfetch leverages Vespa's support for sparse and dense tensors + vector search to do this online in less then 100 milliseconds. 🧵
1
1
19
3,506
Have you ever had to index trillions of documents in your vector database?
This is the problem Yahoo Mail is facing, and believe us it can get expensive.
Luckily this is personal data, so it's possible to use Vespa's Vector Streaming Search which makes the problem tractable.
1
2
15
1,574
The latest Vespa newsletter is here to help you stay up to date on what's happening on the leading edge in RAG, IR and vector search:
- A new SPLADE embedder
- ONNX models with float16
- @cohere embedding model guides
- Support for an array of chunks with ColBERT
- And list of our latest blog posts you shouldn't miss
blog.vespa.ai/vespa-newslett…
3
18
3,188
Very informative post on choosing Vespa for personalization using vector embeddings
Blog post on how @vinted started using @vespaengine for recommendations vinted.engineering/2023/10/0…
2
5
15
2,134
Seems everybody is migrating their search and recommendation systems from Elastic to Vespa now.
Here's the experience of Stanby, Japan's leading job search site: blog.vespa.ai/migrating-to-t…
5
15
1,315
When you are doing nearest neighbor vector retrieval you are doing *search*, which is computation over data.
The llm builders community is currently speedrunning the process of discovering what this takes.
If you are reading this, congrats you're likely ahead of the curve :-)
Vector databases are not search engines.
Reasoning over metadata is easily achievable with COT + @elastic , @vespaengine , or other search engines.
13
4,177
Vector databases are not suited for personal data: You need complete results and much lower cost.
We're announcing vector streaming search to solve this problem: blog.vespa.ai/announcing-vec…
Fully production ready, and open source, of course.
1
5
16
2,149
Benchmark results! Scaling TensorFlow model evaluation with Vespa blog.vespa.ai/post/173669458…
14
15
New blog post: Creating a state of the art question-answering system using Vespa.ai.
blog.vespa.ai/efficient-open…
1
12
16
What we have been up to in the Vespa HQ in August:
🏎️ 30x faster MaxSim with Hamming distance for multivector documents
🧑💻IDE support for VSCode, IntelliJ, PyCharm, WebStorm and neovim
🐍Even more pyVespa improvements
blog.vespa.ai/vespa-newslett…
1
14
1,235
Vector search is more than ranking by nearest neighbor.
Run efficient transformer-based embedding retrieval on your laptop 💻
blog.vespa.ai/accelerating-t…
2
14
1,599
A good writeup on building a search product end to end medium.com/@wabz/building-a-…
"[Vespa] is *the* search engine to use on mutable data sets with modern ranking methods"
4
14
Frozen embeddings 🥶
A guest post on the Vespa blog by Andrii Yurkiv
blog.vespa.ai/leveraging-fro…
1
3
14
4,970
New blog post!
Joining real-time click data and content at query time without impacting performance. Impossible? Not with vespa.ai.
blog.vespa.ai/parent-child-j…
2
14
New vespa.ai blog post: The basics of Vespa applications blog.vespa.ai/post/166978538…
13
14
Getting started with machine-learned ranking using Vespa medium.com/vespa/learning-to…
10
13
Vespa continues to empower developers who are building something real with vector embeddings.
No custom code required to
- use any embedding model from @huggingface
- do multilingual embedding
- run embedders on GPUs
blog.vespa.ai/enhancing-vesp…
1
12
1,348
Is it possible to do distributed joins in real time without sacrificing performance? It depends - is your use case of the parent-child type, and do you run the latest Vespa? blog.vespa.ai/post/174589826…
8
11
We made neural net evaluation 20x faster
blog.vespa.ai/post/169340802…
8
13
What we've been up to during summer at the Vespa.ai hq
- A multi-lingual embedding sample app using E5
- to_epoch_second in the indexing language
- New pyvespa and Vespa CLI features
- ML model persistence across deployments
- and more
blog.vespa.ai/vespa-newslett…
3
12
1,625
Announcing support for ANN with multiple vectors per document in Vespa.ai blog.vespa.ai/semantic-searc…
1
1
9
1,396
Can machines reliably answer questions in natural language?
It turns out they can, and the best methods from research are public.
Seems like a game-changer, so why isn't this more widespread?
1
7
13
How and why Zedge migrated their autosuggest, search and recommendation to Vespa blog.vespa.ai/post/177690706…
4
12
Why you can't conclude a machine-learned model will scale in production just because it is fast blog.vespa.ai/ml-model-servi…
1
4
12
DONE from our ☀️summer☀️ todo list :
- pyVespa deploy_to_prod.
- vespa log CLI command also for self-hosted.
- Faster multithreaded hybrid queries.
- rank-score-drop-limit in second phase.
- Chinese segmentation.
and 14 more, covered in our August issue.
blog.vespa.ai/vespa-newslett…
2
4
12
1,499
Tempted to use a database for your vectors?
Consider why you are using vectors in the first place. blog.vespa.ai/vector-search-…
2
11
1,045
The December Vespa newsletter is out:
- Global-phase ranking with normalizing functions
- Take the Vespa open source survey
- Token access on Vespa Cloud
- Get indexed tokens in results
- and much more
blog.vespa.ai/vespa-newslett…
11
2,848
Vespa improvements released in February: Support for LightGBM machine-learned models, improved matrix multiplication performance, benchmarking guide, a fluent query builder API, and Hadoop integration improvements. blog.vespa.ai/vespa-product-…
6
12
Vespa improvements from January blog.vespa.ai/vespa-product-…
Among other things this includes new tensor functions needed to run BERT models. All of it already released and running in production of course.
1
7
12
Highlights of Vespa features released in June:
- Personal vector search with complete results, at 1/20 of the cost, with streaming
- MIPS/dot product ranking in ANN
- GPU Acceleration of embedding models
- Use any @huggingface embedder directly
blog.vespa.ai/vespa-newslett…
3
10
2,544
Everybody knows that AI builds on tensors now, but when we started work on them in 2013 it was just a weird thing.
1
3
12
If you are looking for a comparison of Elasticsearch, Solr, and Vespa this is the video to watch.
Find out how our invited speakers compared #Elasticsearch #Solr and @vespaengine at the Haystack LIVE! Meetup - video now available at piped.video/watch?v=SzZ_A9G6… - thanks so much @anshumgupta @jobergum @joshdevins !
2
11
If you are using vector embeddings, reading this post might be the most profitable ten minutes you'll ever spend.
Matryoshka 🤝 Binary vectors: Slash vector search costs with Vespa
We announce support for combining matryoshka and binary quantization in Vespa’s native hugging-face embedder and discuss how this slashes vector search costs.
blog.vespa.ai/combining-matr…
1
10
1,440
Implementing reinforcement learning in production with open source: New Vespa blog post medium.com/vespa/serving-art…
9
11
Gigaom published their Sonar for Vector Databases today, positioning Vespa as a leader.
While what we are - and what you need - is much more than a vector database, it is gratifying to be recognized as a leader also on these core features alone.
1
2
11
948
Finally it is possible to create search systems that can see: Drop document text extraction and let your system see text, figures and layout directly.
blog.vespa.ai/scaling-colpal…
2
11
960
Announcing GPU-accelerated inference in Vespa blog.vespa.ai/gpu-accelerate…
1
2
10
961
It's been a good summer for coding in the north 🌧️🌧️
We got some new features in Vespa released: BM25 ranking feature, searchable parents, tensor summary features and metric export
blog.vespa.ai/post/187148684…
4
11
Scaling vector search to a billion documents per node blog.vespa.ai/billion-scale-…
1
10
It turns out that when you open source 1.7M lines of code in 150 flat modules people will keep asking for directions. Today we're publishing a map to the Vespa.ai code base.
github.com/vespa-engine/vesp…
3
11
April @vespaengine updates include performance and operability improvements: Top-K hits, smarter data migration and CloudWatch integration. Contributing to Vespa is now easier with the release of a CentOS 7 dev environment. blog.vespa.ai/vespa-product-… #bigdata
6
11
When GigaOm named Vespa Leader in their Sonar for Vector Databases, one of the categories where we scored Excellent were Embedding Flexibility - why?
Vespa lets you create embeddings in four ways:
- On your own, outside Vespa: Just pass tensors directly in documents and queries.
- With your own model, run by Vespa: Add the model to the application package and reference it in embed expressions.
- With a model provided by Vespa: Reference a model on the Vespa Cloud model hub in your embed expression.
- With custom code doing what you want in a custom Embedder, or - if you want full control over the process - a custom Searcher and Docproc.
These also allow you to create and index a collection of embeddings for a single document, either for each token of the text, for an array of chunks, or both at the same time!
You can use any of these methods at the same time for different fields, and change at any time without changing any other aspect of the application. This lets you get started easily with embeddings while also empowering you to add more sophisticated methods gradually. (The golden combination we strive for in all things).
2
11
1,274
💥New blog post: Improving Zero-Shot Ranking with Vespa Hybrid Search
blog.vespa.ai/improving-zero…
1
9
1,096
The big data maturity levels medium.com/@bratseth/the-big…
Tag your org!
1
4
10
When you're working with vectors you're doing search - and you'll need all the features of a search engine.
A new blog post which explains why we're no fans of the term 'vector database'
blog.vespa.ai/when-you-are-u…
2
10
856
Semantic search with multiple languages is finally clearing the bar for production.
Vespa makes it easy to stand up your own.
blog.vespa.ai/simplify-searc…
2
9
908
We're submitting some Vespa examples to relevance competitions. All of them:
- available as open source sample applications
- production ready, with < 100 ms latency
This first one is a simple WAND+GDBT baseline, beating all the other entries not using deep learning/embeddings.
Welcome @jobergum and @vespaengine to the fray with their first @MSMarcoAI document leaderboard submission! Awesome to see LTR effectiveness creeping up on the muppets. microsoft.github.io/MSMARCO-…
1
3
10
How would you implement the next Gmail efficiently in open source? New Vespa blog post medium.com/vespa/efficient-p…
7
9
Did you know you can now convert text to vector embeddings in documents and queries automatically in Vespa? buff.ly/3CnNqlj
We provide the wonderful SentencePiece algorithm by Taku@Google for this, but you can also plug in your own.
2
10
And how to use a BERT-style transformer model for final ranking while staying within your latency budget blog.vespa.ai/pretrained-tra…
2
10
Finally, for all the Elastic developers out there who want to start creating modern search base applications, @atitaarora on understanding Vespa with a Lucene mindset piped.video/_ML-QB0Zxvg
3
9
For those who are running Vespa, information about HTTP/2 Rapid Reset (CVE-2023-44487)
blog.vespa.ai/cve-2023-44487…
1
9
1,528
Testing ANN performance on vespa.ai: Search a million vectors in 2 ms measured from the client.
From our performance factory page at @vespaengine HQ where we test Vespa performance for every build. This is testing Vespa's approximate nearest neighbor search on SIFT 1M blog.vespa.ai/approximate-ne…
This is an end to end benchmark including HTTP API. Pretty awesome if you ask me.
3
9
ColBERT v2 is the state of the art in information retrieval, and with this work you can actually run it economically, at any scale.
Announcing ColBERT in @vespaengine, enjoy!
- A new native Vespa ColBERT v2 embedder
- ColBERT token-level vector compression (32x)
- Support for long context via Vespa mixed tensors
- Offload to disk
- Eval
Plus, it boasts the largest FAQ ever!😅
blog.vespa.ai/announcing-col…
9
1,042
October news from Vespa.ai 🎃
🌤️ Enclave: Bring your own cloud to Vespa Cloud, on both AWS and GCP.
🏎️MUCH faster fuzzy matching.
🔍Lucene linguistics integration.
... and much more obviously: blog.vespa.ai/vespa-newslett…
3
9
1,393
This lets you combine vector search with filters and text while staying efficient, and scale to real data sizes with low latency.
blog.vespa.ai/using-approxim…
This is what's needed for production and what can't be achieved by integrating specialist tools for these problems.
[3/4]
1
8
If you're in Bay Area come meet us at September 26. in San Francisco meetup.com/SF-Big-Analytics/… Thanks to @Amplitude_HQ for hosting!
5
9
If you have Vespa.ai it's not too late to make yourself a state of the art and infinitely scalable e-commerce site in time for Black Friday.
@jobergum explains how in medium.com/vespa/e-commerce-…
9
9
For more on how to use reinforcement learning for recommendation see medium.com/vespa/serving-art…
A breath of fresh air: @vespaengine at #HaystackConf, signal to noise is through the roof! Reinforcement learning for comment ranking.
6
6
Announcing a new IN operator in Vespa:
select * from product where id in (10, 20, 30)
- Simpler than using a WeightedSet
- Over 10x faster with large thousands of query values
blog.vespa.ai/announcing-in-…
1
8
740
ANN lets us apply AI in ways that weren't possible before.
But to use it in the real world you need a few more things.
blog.vespa.ai/using-approxim…
1
8
Shipped from the Vespa HQ in 🎃 October:
- Index tensors with multiple sparse dimensions: tensor(page{}, section{}, chunk{}, x[16])
- Global significance ("tf") models- Metric dashboards in the Vespa Cloud Console
- Even more new PyVespa features
Read the details in our latest newsletter: blog.vespa.ai/vespa-newslett…
1
7
718
Will new vector databases dislodge traditional search engines?
blog.vespa.ai/will-vector-db…
2
8
Want to run Vespa.ai on your m1 Mac, on ARM64 nodes in your data centers, or on Vespa Cloud?
We got you! All Vespa images are now multi-architecture.
blog.vespa.ai/vespa-on-arm64…
2
7
Scalable realtime blog recommendation using neural nets
blog.vespa.ai/post/168572816…
Part 3 of 3 of a series showing in detail how to build a big data serving application using Vespa.
6
8
An open source sample application for searching images by describing them, based on CLIP buff.ly/3IvNG4V
3
7
We also provide the same for *your* applications when running on Vespa Cloud.
@vespaengine has one of the best FOSS ci/cd set up I have seen out there. Pretty much everything is automated. This encourages contributors like me to contribute even more because the changes are so instant on prod.
7
795
Danswer (YC W24) on why they moved their RAG solution to Vespa.
We hear this in so many private conversations right now.
@DanswerAI has been using @vespaengine as our search engine for a long time. In this blog, I outline the key benefits of using Vespa.
Thanks @jonbratseth, @kraune and @jobergum for keeping Vespa open source and helping our users with Vespa questions!
blog.vespa.ai/why-danswer-us…
7
3,270
For those who want to contribute to Vespa.ai this will be an easy way to get started
Join us March 21 - 28, for Yahoo Hack Together, a virtual #opensource #hackathon! Learn more & register: yahoo.com/hacktogether. #bigdata #design #devops #networksecurity
3
7
The full post includes a detailed description of how to implement a scalable and low latency recommender system: medium.com/farfetch-tech-blo…
7
498
Vespa product updates for December:
- 10-150x speedup of computations with sparse tensor dimensions
- ZooKeeper distributed locking available in containers
- PyVespa for machine learning
- ONNX runtime integration
blog.vespa.ai/vespa-product-…
2
7
🎅The official 2024 Advent of Tensors 🎅
Follow this thread to get a challenge a day, win swag and become a tensor computation expert along the way.
Prepare to embark on a festive journey as we bring you the Advent of Tensors with 24 challenges and the chance to win @vespaengine swag!
blog.vespa.ai/advent-of-tens…
3
7
1,059
This is an unusually good comparison, although it leaves out what we see as most important: Ranking and inference capabilities.
It can be solved elsewhere, but only if you don't need to perform at scale ...
Found this comprehensive spreadsheet on the key features of popular Vector Databases used to build out AI chatbot solutions.
Spreadsheet: t.ly/FZxT4
Key takeaways:
- The most well-rounded solutions include @weaviate_io, @vespaengine, and @elastic.
- The top open-source solutions include @qdrant_engine, Weaviate, PG Vector, Vespa, @trychroma, and @milvusio.
- Weaviate and Vespa allow you to integrate your own embeddings model, which you may have optimized for your use case.
- Most Vector Databases support metadata filtering, which enables you to narrow the scope of the embeddings search and improve accuracy.
- @pinecone has the lowest metadata size limit (40kb), whilst Qdrant and Vespa provide unlimited size limits.
- Most Vector Databases provide integrations with popular AI development tools @langchain and @llama_index.
Credit: t.ly/4irYt
7
865