Weaviate Podcast

Hey everyone, I am SUPER excited to present our 76th Weaviate Podcast featuring Patrick Lewis, an NLP Research Scientist at Cohere! Patrick has had an absolutely massive impact on Natural Language Processing with AI and Deep Learning! Especially notable for the current climate in AI and Weaviate is that Patrick is the lead author of the original "Retrieval-Augmented Generation" paper!! Patrick has contributed to many other profoundly impactful papers in the space as well such as DPR, Atlas, Task-Aware Retrieval with Instruction, and many many others! This was such an illuminating conversation, here is a quick overview of the chapters in the podcast! 1. Origin of RAG - Patrick explains the build-up that lead to the RAG paper, AskJeeves, IBM Watson, conceptual shift to retrieve-read in mainstream connectionist approaches to AI. 2. Atlas - Atlas shows that a much smaller LLM when paired with Retrieval-Augmentation can still achieve competitive few-shot and zero-shot task performance. This is super impactful because this few-shot and zero-shot capability has been a massive evangelist for AI broadly, and the fact that smaller Retrieval-Augmented models can do this is massive for the economically unlocking these applications. Teasing apart some architectural details of RAG: 3. Fusion In-Decoder - Interesting encoder-decoder transformer design in which each document + the query is encoded separately, then concatenated and passed to the LM. 4. End-to-End RAG - How to think about jointly training an embedding model and an LLM augmented with retrieval? 5. Query Routers - How to route queries from say SQL or Vector DBs? (More nuance on this later with Multi-Index Retrieval) 6. ConcurrentQA - Super interesting work on the privacy of multi-index routers. For example, if you ask "Who is the father of our new CEO" - this may reveal the private information of the new CEO with the public query of their father. 7. Multi-Index Retrieval 8. New APIs for LLMs 9. Self-Instructed Gorillas 10. Task-Aware Retrieval with Instructions 11. Editing Text, EditEval and PEER 12. What future direction excites you the most? Links: Learn more about Patrick Lewis: https://www.patricklewis.io/ Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks: https://arxiv.org/abs/2005.11401 Atlas: https://arxiv.org/pdf/2208.03299.pdf Fusion In-Decoder: https://arxiv.org/pdf/2007.01282.pdf Chapters 0:00 Welcome Patrick Lewis! 0:36 Origin of RAG 5:20 Atlas 10:43 Fusion In-Decoder 17:50 End-to-End RAG 27:05 Query Routers 32:05 ConcurrentQA 37:30 Multi-Index Retrieval 40:05 New APIs for LLMs 41:50 Self-Instructed Gorillas 44:35 Task-Aware Retrieval with Instructions 52:00 Editing Text, EditEval and PEER 55:35 What future direction excites you the most?

Hey everyone! Thank you so much for watching the 75th Weaviate Podcast featuring Tanmay Chopra! The podcast details Tanmay's incredible career in Machine Learning from Tik Tok to Neeva and now building his own startup, Emissary! Tanmay shared some amazing insights into Search AI such as how to process Temporal Queries, how to think about diversity in Retrieval, and Query Recommendation products! We then dove into the opportunity Tanmay sees in fine-tuning LLMs and knowledge distillation that motivated Tanmay to build Emissary! I thought Tanmay's analogy of GPT-4 to 3D printers was really interesting, tons of great nuggets in here! I really hope you enjoy the podcast, as always more than happy to answer any questions or discuss any ideas with you related to the content in the podcast! Chapters 0:00 Welcome Tanmay! 0:23 Early Career Story 2:02 Tik Tok 4:10 Neeva 8:45 Temporal Queries 11:40 Retrieval Diversity 17:22 Query Recommendation 23:20 Emissary, starting a company! 30:20 A Simple API for Custom Models 35:42 GPT-4 = 3D Printer?

Hey everyone! Thank you so much for watching the 74th Weaviate Podcast feature Simba Khadder, the CEO and Co-Founder of FeatureForm! To begin, "features" broadly describe the inputs to machine learning models that they use to produce outputs, or predictions. Feature stores orchestrate the construction of features, whether that be transformations for tabular machine learning models such as XGBoost, to chunking for vector embedding inference, and now features for LLM inference in RAG. Right out of the gate, Simba really opened my eyes to the role that feature engineering plays in RAG. Further touching on this at the very end under the "Exciting future for RAG with Features" chapter, Simba further describes how we can use more advanced features to provide better context to LLMs. In addition to these insights on RAG, there are so many nuggets in the podcast, Simba is a world class professional when it comes to building distributed systems, production scale recommendation systems, and more! I learned so much from chatting with Simba, I hope you enjoy listening to the podcast! As always we are more than happy to answer any questions or discuss any ideas you have about the content in the podcast! FeatureForm: https://www.featureform.com/ Highly Recommend!! Simba Khadder at the CMU DB Seminar series: https://www.youtube.com/watch?v=ZsWa6XiBc-U FeatureForm and Weaviate demo! https://docs.featureform.com/providers/weaviate Chapters 0:00 Simba Khadder 0:35 RAG and Feature Stores 4:30 Experience building Recommendation Systems 9:47 The End-to-End Feature Lifecycle 15:08 Virtual Feature Store Orchestration 26:45 RAG Evaluation 31:27 Feature Engineering 34:15 LLM Tuning and Features 39:55 Streaming Features 51:15 Data Drift Detection 54:20 Exciting future for RAG with Features

Charles Packer, lead author of MemGPT at UC Berkeley, discusses the concept of explicit memory management in GPT models, the use of prompts to handle memory limitations, interrupts in retrieval augmented generation (RAG), achieving ideal running speed in high parameter models, fine-tuning MemGBT for long conversations, search actions pagination, role-playing language models, and the future integration of memory in chatbot platforms.

Hey everyone! Thank you so much for watching the 72nd episode of the Weaviate Podcast with Madelon Hulsebos!! Madelon is one of the world's experts on Machine Learning with Tables and Tabular-Structured Data, this was such an eye-opening conversation! We discussed all sorts of topics from the relationship of tabular data and embeddings, to searching through tables, semantic joins, more complex Text-to-SQL, using machine learning for query execution, using tabular data in search and recommendation reranking, and many more! This was easily one of the most knowledge packed episodes of the Weaviate podcast so far, please don't hesitate to leave any questions or ideas you have related to the content discussed! You can learn more about Madelon's incredible research career and publications / talks here: https://www.madelonhulsebos.com/! Papers such as GitTables are listed here! Another nice nugget form the podcast - Madelon introduced me to the BIRD-SQL benchmark which really expanded my understanding of Text-to-SQL (https://arxiv.org/pdf/2305.03111.pdf. Chapters 0:00 Welcome Madelon! 0:58 Tabular Data and Embeddings 3:10 Tabular Representation Learning 5:48 Semantic Type Detection 9:50 Pandas as an LLM Tool 11:52 Table-Based Question Answering and Text-to-SQL 19:35 Joins with Machine Learning 21:38 Query Execution with Machine Learning 22:45 Graph Neural Networks 24:07 XGBoost 28:28 Merging Tables 32:10 Fact Representation 35:50 GPT-4V and Tables 39:00 Metadata in Embeddings 42:45 Table Retrieval in Weaviate 46:25 Exciting future directions!!

Hey everyone! Thank you so much for watching the 71st Weaviate Podcast with Vibs Abhishek! Vibs is the CEO and Founder of Alltius AI, as well as a professor at UC Irvine business school! In order to tame the somewhat chaotic emerging landscape of RAG and LLM applications, Alltius has settled on 3 core pillars of Knowledge, Skills, and Deployment Channels! Vibs further explained how he sees the distinction between Assistants and Agents and many more topics important to Enterprise deployment of RAG applications such as reducing hallucinations and employing classifiers to route skills and knowledge sources! I learned so much from this conversation, I hope you enjoy the podcast! Alltius KNO Plus Demo Video: https://www.loom.com/share/fcfe516b75ea4f069b1a8d6a3510fa4c?sid=5f43317f-c20b-4dd9-91d3-2cde993fd91f Chapters 0:00 Welcome Vibs 0:22 Background 2:30 Alltius’ UI for Assistants 7:15 The Knowledge Pillar 12:05 SQL Router and Intent Management 14:10 Classifying a Pipeline / Skill 17:30 Flexibility of Zero-Shot versus Fine-Tuning 21:00 The Channels Pillar 23:00 Connecting the Warehouse / Lakehouse 24:50 Assistant versus Agent 28:30 MemGPT 31:25 Offline LLM Research 35:50 Multi-Agent Role-Playing Assistants 39:25 From Clicks to Conversations 44:10 CEO / Professor and Evolution of the Field

Discover the innovative world of MemGPT, where operating system principles meet large language models. Explore how memory management is revolutionized to enhance conversational AI. Delve into the architecture that boosts dialogue consistency and engagement. Unpack the challenges of training long-context models and the role of efficient memory in search dynamics. Learn about the creation of synthetic textbooks as training data, showcasing the seamless interaction of language models and APIs.

Hey everyone! Thank you so much for watching the 70th episode of the Weaviate podcast with Neum AI CTO and Co-Founder Kevin Cohen! I first met Kevin when he was debugging an issue with his distributed node utilization and have since learned so much from him about how he sees the space of Data Ingestion, also commonly referenced as ETL for LLMs! There are so many interesting parts to this from the general flow of data connectors, chunkers and metadata extractors, embedding inference, and the last leg of the mile of importing the vectors to a Vector DB such as Weaviate! I really loved how Kevin broke down the distributed messaging queue and system design for orchestrating data ingestion at massive scale such as dealing with failures and optimizing the infrastructure as code setup. We also discussed things like new use cases with quadrillion scale vector indexes and the role of knowledge graphs in all this! I really hope you enjoy the podcast, please check out this amazing article below from Neum AI! https://medium.com/@neum_ai/retrieval-augmented-generation-at-scale-building-a-distributed-system-for-synchronizing-and-eaa29162521 Chapters 0:00 Check this out! 1:18 Welcome Kevin! 1:58 Founding Neum AI 6:55 Data Ingestion, End-to-End Overview 9:10 Chunking and Metadata Extraction 14:20 Embedding Cache 16:57 Distributed Messaging Queues 22:15 Embeddings Cache ELI5 25:30 Customizing Weaviate Kubernetes 38:10 Multi-Tenancy and Resource Allocation 39:20 Billion-Scale Vector Search 45:05 Knowledge Graphs 52:10 Y Combinator Experience

Hey everyone! Thank you so much for watching the 69th episode of the Weaviate Podcast featuring Charles Pierse from Tactic! Tactic has recently launched their new Tactic Generate project, an incredible UI for conducting research across multiple documents. I think there is a massive opportunity to pair these prompts and LLM workflows with User Interfaces and take more of a holistic User Experience perspective. Tactic Generate has done an incredible job of that, please take a look from the link below! I had such a fun conversation catching up with Charles (Charles was our 2nd Weaviate Podcast guest!), I hope you enjoy the podcast! Tactic Generate: https://tactic.fyi/generative-insights/ Chapters 0:00 Tactic Generate 1:40 Welcome Charles! 2:38 Charles’ work at Tactic 4:40 LLMs comparing documents 9:10 LLM Chaining 17:30 Discovering LLM Chains 20:28 Moats in ML Products 28:48 Fine-Tuning vs. RAG 34:30 Fine-Tuning Search Models 39:45 Skepticism on RLHF 41:52 Gorilla, Integrations, and CRM 45:40 Query Routers 47:55 CRM and Tree-of-Thoughts 55:54 Graph Embeddings 1:02:20 Llama CPP / GGML 1:04:28 What are you looking forward to most in AI?

Hey everyone! Thank you so much for watching the 68th episode of the Weaviate Podcast! We are super excited to welcome Morgan McGuire, Darek Kleczek, and Thomas Capelle! This was such a fun discussion beginning with generally how see the space of fine-tuning from why you would want to do it, to the available tooling, intersection with RAG and more! Check out W&B Prompts! https://wandb.ai/site/prompts Check out the W&B Tiny Llama Report! https://wandb.ai/capecape/llamac/reports/Training-Tiny-Llamas-for-Fun-and-Science--Vmlldzo1MDM2MDg0 Chapters 0:00 Tiny Llamas! 1:53 Welcome! 2:22 LLM Fine-Tuning 5:25 Tooling for Fine-Tuning 7:55 Why Fine-Tune? 9:55 RAG vs. Fine-Tuning 12:25 Knowledge Distillation 14:40 Gorilla LLMs 18:25 Open-Source LLMs 22:48 Jonathan Frankle on W&B 23:45 Data Quality for LLM Training 25:55 W&B for Data Versioning 27:25 Curriculum Learning 29:28 GPU Rich and Data Quality 30:30 Vector DBs and Data Quality 32:50 Tuning Training with Weights & Biases 35:47 Training Reports 42:28 HF Collections and W&B Sweeps 44:50 Exciting Directions for AI