Weaviate Podcast

Weaviate podcast #33. Thank you so much for watching the 33rd Weaviate Podcast! This episode features one of the heroes of Deep Learning for Search, Nils Reimers! Nils' work on SentenceBERT is one of the foundational works for applying Deep Representation Learning to text search. This is the idea that personally inspired me to work in this field. Having seen the successes of Contrastive Representation Learning for Computer Vision, I was mind-blown by the possibility of this for NLP and text search. In addition to the scientific foundation, the software development of the Sentence Transformers library and BEIR benchmarks has been enormously impactful! It was an honor getting to ask Nils the questions I have about these things, from the role of Data Quality to Intent, Sparse Vectors, Long Document Encoding, Distribution Shift, and many more. I really hope you enjoy the podcast! We are so excited about the Cohere Multilingual embedding model and can't wait to see what else comes out of Cohere and their amazing team! Cohere Multilingual ML Models with Weaviate: https://weaviate.io/blog/2022/12/Cohe... Nils Reimers: https://scholar.google.com/citations?... Mentioned in the podcast, Cross-Encoders: https://weaviate.io/blog/2022/08/Usin... How to choose a Sentence Transformer from HuggingFace: https://weaviate.io/blog/2022/10/How-... Chapters 0:00 Cohere X Weaviate 0:22 Welcome Nils Reimers! 1:18 Origin Story 3:15 Learning Text Embeddings 6:54 Positive and Negative Sampling in Contrastive Learning 13:32 1 Billion Pairs for Text Embedding Optimization 15:44 Impact of Data Quality 18:40 New Cohere Multilingual Model! 24:50 Challenge of Debugging Multilingual Models 28:30 Intent in Search 30:40 Thoughts on ColBERT 33:50 Sparse Vectors in Search 36:17 Long Documents and Multi-Discourse 43:40 Entity Parsing in Query Understanding 46:08 Unknown Words and Distribution Shift 50:07 Re-Vectorizing with Fine-Tuning 53:07 More on Search Interfaces and Intent in Search 55:15 Thank you Nils!

Weaviate Podcast #32. Thank you so much for watching the Weaviate podcast! We are super excited to host Sam Bean from You.com! As well as welcome Zain Hasan and John Trengrove to the Weaviate podcast for the first time! Sam begins by describing You.com, and then we dive into the Weaviate Spark Connector that Sam played a massive role in creating. I thought this was such a masterclass in the Spark big data technology; John, Sam, and Zain are all data engineering pros and I've never learned more about a new technology from a podcast than this one. I really hope you enjoy listening to it, please let us know any questions or ideas you have. Also, please see Zain's blog post on "The Details Behind the Sphere Dataset in Weaviate" - https://weaviate.io/blog/2022/12/deta.... This provides great detail on exactly how to use the Spark connector in Weaviate! In this case for a billion-scale dataset upload!!! Chapters 0:00 Thanks for Watching! 0:18 Welcome Zain and John 0:28 Welcome Sam Bean, You.com! 1:48 Search Interface and Search Apps / Widgets 3:40 Searching through Specific Websites 4:00 Origin Story of You.com 6:53 How did you come across Weaviate? 8:33 Text, Image, Audio Search 10:28 What do you use Spark for? 14:20 Datasets used with Weaviate 16:14 Creating a Spark Connector to Weaviate 21:05 Adding Streaming support 22:50 Vectorizing Data at You.com 27:15 More on ONNX + Spark 29:52 Performance Questions, Spark + Weaviate 34:35 Parquet for HuggingFace Dataset Files 34:54 What is Parquet? Spark Pushdown Filters Explained 39:04 Similar to HDF5? 39:45 Spark for Extracting Ranking Features 43:25 Hybrid Search 46:48 Collecting Search Relevance Data 51:07 Thank you for watching! Thanks Sam, Zain, and John!

Weaviate Podcast #31.  Weaviate 1.17!! This is a massive release for Weaviate, debuting Replication, Hybrid Search, BM25, Faster Startup and Import Times, as well as other fixes! Replication and Hybrid Search are two massive features for Weaviate, and we really hope you enjoy the description of them from the podcast. Please also check out the Weaviate 1.17 release blog post for more information as well - https://weaviate.io/blog/2022/12/Weaviate-release-1-17.html!   This is also a very special podcast as we welcome Parker Duckworth for the first time to the podcast! Parker gave an excellent explanation of Replication and unpacked some of the questions we are seeing around Ref2Vec!  Thank you so much for listening to the podcast! Please check out the newest version of Weaviate!   Chapters  0:00 Weaviate 1.17! Welcome Parker!  0:28 From Italy to 1.17  2:04 Replication work in Italy  3:58 Replication Details  6:28 Use Cases of Replication  13:12 Product Engineering  16:24 Hybrid Search  21:30 Open Question around Hybrid Search  23:15 Rank Fusion 24:00 BEIR Benchmarks  27:28 What is Ref2Vec?  29:08 Bipartite Graph Ref2Vec Example  29:30 Graphs in Weaviate  34:25 Ref2Vec Cascading Updates  37:45 Custom Aggregation Functions in Ref2Vec  39:08 Adding Recency Bias in Ref2Vec  41:18 Startup Time Improvements  41:50 Batch Latency Improvement

Weaviate Podcast #30.  Chapters 0:00 The future of search! 0:42 Welcome Marco and Chris! 4:28 Solving Hallucination with External Memory LLMs 8:16 Bob van Luijt on Weaviate and LLMs, Collaborations 14:48 What we have is not yet what the technology is capable of 16:45 Everything is Search! 18:55 The Magic of Machine Learning 20:30 Asking follow up questions 22:28 Meaning in LLMs and RLHF 27:10 How ChatGPT is Evangelizing the Technology 29:45 What is the future of search from a user perspective? 34:38 Integration with Existing Businesses 35:20 Impact on Creativity 37:37 Data Visualization from Natural Language Questions 39:00 Thought experiment - is this notebook an extension of the brain? 42:20 More on “always on” interface 43:42 General vs. Specific Intelligence 45:25 Software Business Impact 48:12 Open-Source Models 49:25 Finding Niches 56:30 Pride in Humans + AI 57:20 Exploring more Prompts 59:20 Personalized Embedding Key 1:02:40 Concluding Thoughts

Weaviate Podcast #29. Hey everyone, thank you so much for watching another episode of the Weaviate podcast! This episode features Matthijs Douze, one of the most talented and accomplished scientists we've hosted on the Weaviate podcast! Matthijs has pioneered the use of Product Quantization to compress vector representations and enable even faster and more efficient approximate nearest neighbor vector search. Matthijs told an incredible story about the history of this research, from searching from SIFT vectors for Computer Vision Search applications like real-time CD Cover album search to the problems facing modern IVF-PQ systems and the use of PQ in graph-based HNSW search. This is also a very special episode as Abdel Rodriguez makes his debut on the Weaviate podcast to discuss Weaviate's efforts in integrating PQ support and the unique challenges with this algorithm and the incremental updates required for a Vector Database. On this topic, Etienne Dilocker also returned to discuss the topic of Vector Database vs. Library with Matthijs, who is one of the lead developers of the Faiss library. This was a really information-heavy podcast, please don't hesitate to ask us any questions or present any of your ideas! Thanks again for listening!

Weaviate Podcast #28. Thank you so much for watching the 28th Weaviate Podcast! This episode features Maarten Grootendorst, developer of the BERTopic python library and an active evangelist of this exciting cluster analysis technology, (Maarten has written some incredible articles here - https://medium.com/@maartengrootendorst)! In this podcast, Maarten did an incredible job explaining how BERTopic works, with particular details such as k-Means clustering vs. HDBSCAN, Semi-Supervised topic modeling, Dynamic topic modeling, and many more! I was amazed at Maarten's expertise in the miscellaneous details of these algorithms! We are extremely excited about adding BERTopic to Weaviate, please see this proposal if interested in contributing to the discussion: https://github.com/semi-technologies/...!

Weaviate Podcast #27. Thank you so much for watching the 27th episode of the Weaviate Podcast! This is truly one of my favorite podcasts we have published so far, I think the way Weaviate and Neural Magic fit together is really exciting! Michael did an amazing job explaining the concepts behind how Neural Magic delivers and tests inference acceleration, as well as the vision for the future of Deep Learning with Sparsity and CPU inference. I really hope you enjoy the podcast, more than happy to answer any questions or entertain any ideas/discussion! Thanks again for watching! Weaviate users can begin using Neural Magic's text vectorization pipeline as a custom text2vec-transformers docker image here - cshorten/experimental-text2vec-neuralmagic. Please note this is an experimental build and we will be releasing the full integration with thorough testing very soon!

Weaviate Podcast #26. Thank you so much for watching the 26th episode of the Weaviate Podcast! This is another really special episode! Jonathan Frankle is one of the world's experts in Deep Learning and is making incredible advances at MosaicML in efficient Deep Learning training. The headline event is the release of MosaicML Cloud and a set of new cost estimates for GPT language models at different scales (linked below). Jonathan explains that these numbers are a baseline and he predicts they could get to as low as $100K as they seek opportunities for efficiency optimizations. This story has already played out in the realm of ResNet ImageNet training as MosaicML has demolished expectations of how fast we can train these models and it seems highly likely they will do the same for large language model costs. Jonathan and I also discussed the general space of Language Models and their applications, especially discussing their role as Databases in things like the Weaviate Vector Search Engine. We also discussed Self-Ask, Chain-of-thought Prompting, and tool use in Language Models. I had an awesome time picking Jonathan's brain about these topics and I hope you all enjoy the podcast, more than happy to answer any questions or entertain any ideas / discussion! Thanks again for watching!  Blog post: GPT-3 Quality for less than $500K - https://www.mosaicml.com/blog/gpt-3-q...

Weaviate Podcast #25. Thank you so much for watching the 25th episode of the Weaviate Podcast! This is a really special episode with Erik Bernhardsson! Erik is one of the early thought leaders on Approximate Nearest Neighbor (ANN) Search, creating the ANNOY library at Spotify. Erik shared incredible insights about vector search at Spotify such as the role of Offline and Online Machine Learning inference and the role of multi-stage re-ranking pipelines. Erik has also done massively impactful work on benchmarking ANN algorithms! We really hope you enjoy the podcast and would be thrilled to answer any questions you have about the conversation topics! Thanks again for watching!

Weaviate Podcast #24. Weaviate v1.15 Release! Thank you so much for checking out the Weaviate podcast -- here is a summary of what is new in Weaviate 1.15:  1. Cloud-native backups – allows you to configure your environment to create backups – of selected classes or the whole database – straight into AWS S3, GCS or local filesystem 2. Reduced memory usage - we found new ways to optimize memory usage, reducing RAM usage by 10-30%. 3. Better control over Garbage Collector – with the introduction of GOMEMLIMIT we gained more control over the garbage collector, which significantly reduced the chances of OOM kills for your Weaviate setups. 4. Faster imports for ordered data – by extending the Binary Search Tree structure with a self-balancing Red-black tree, we were able to speed up imports from O(n) to O(log n) 5. More efficient filtered aggregations – thanks to optimization to a library reading binary data, filtered aggregations are now 10-20 faster and require a lot less memory. 6. Two new distance metrics – with the addition of Hamming and Manhattan distance metrics, you can choose the metric (or a combination of) to best suit your data and use case. 7. Two new Weaviate modules – with the Summarization module, you can summarize any text on the fly, while with the HuggingFace module, you can use compatible transformers from the HuggingFace 8. Other improvements and bug fixes – it goes without saying that with every Weaviate release, we strive to make Weaviate more stable – through bug fixes – and more efficient – through many optimizations.  Please check out this awesome blog post from Sebastian Witalec and the team describing these further - https://weaviate.io/blog/2022/09/Weav....