Vector Podcast

Vector Podcast

di Dmitry Kan
Miglioramento personaleScienzaTecnologia
Stagione 4
Berlin Buzzwords 2026 - Trey Grainger & Doug Turnbull, Role of Search in modern AI and new course
This episode was recorded LIVE at the Berlin Buzzwords 2026 YouTube version: https://youtu.be/acOGVynTVpM The cameraman is pure Zoom's AI ;) The Course: "AI-Powered Search: Modern Retrieval for Humans & Agents" aipoweredsearch.com/live-course?promoCode=vector-podcast Discount Code for course (20% off): "vector-podcast" AI-Powered Search (Book, Content, Community): https://aipoweredsearch.com/ Timecodes 00:00 Intro 00:30 Doug's and Trey's impression of the conference 01:16 How modern AI changed search (if any) 05:48 How to bring AI techniques into existing search engines on a budget 10:44 What the AI-Powered Search course includes 18:02 Staying hands-on 18:47 Guest's favourite topic that keeps them up at night 27:40 Message to the builders of search tech 32:55 Search continues to be challenging and exciting Shownotes: - Upcoming course in detail (grab promo code above to save 20%): https://aipoweredsearch.com/articles/the-frontier-of-ai-search-ai-powered-search-modern-retrieval-for-humans-agents/ - Doug's blog on search, agents, RAG, LLM as a judge and more: https://softwaredoug.com/ - AI-Powered Search: https://aipoweredsearch.com/ - Berlin Buzzwords: - MICES: https://mices.co/ - Future of Search conference: https://berlinsearchweek.com/future-of-search/ - Women of Search: https://www.women-of-search.org/ - "AI is here - time to throw away our search engines?" panel hosted by Charlie Hull at Berlin Buzzwords 2026: https://www.youtube.com/watch?v=StaPk0k-52Y - Dmitry's prototype of Wormhole vectors idea with OpenSearch: https://aiven.io/blog/beyond-hybrid-search-traversing-vector-spaces-with-wormhole-vectors - Dmitry's blog on Medium: https://dmitry-kan.medium.com/ - Dmitry's Tech Stories on Substack: https://substack.com/@dmitrykan - Follow me on LinkedIn for Search updates: https://www.linkedin.com/in/dmitrykan/
S4 · E5
21 giu 2026
39:21
Beyond Hyperspace - Ohad Levi on Hardware Accelerated Search and Agentic Memory
In this episode we sat down with Ohad Levi, co-founder and CEO of Hyperspace, to discuss the harware-accelerated search product he has built to address the search latency problem. Ohad also shares his thoughts on Agentic memory and what keeps him at night these days. Podcast design by https://www.linkedin.com/in/srbhr/ Timecodes: 00:00 Intro 01:35 Ohad's background 03:30 How idea was born: what was missing in the search landscape 06:52 Top 3 issues with existing search solutions 10:52 The importance of search latency 13:41 Ohad's solution for latency 19:22 Was Hyperspace up for the challenge? 22:12 New approaches to handling massive scale 26:12 Does latency matter for new agentic AI? 32:12 Agentic AI vs SaaS 35:03 Ohad's learnings from Hyperspace 38:37 Friction points for the hardware-accelerated search 42:40 Product-led growth way 47:43 What keeps Ohad excited about the AI / search field 51:43 Ohad's message to the Search community Shownotes: Ohad Levi on LinkedIn: https://www.linkedin.com/in/ohad-levi/ Hyperspace: https://www.hyper-space.io/ Dmitry's blog on Medium: https://dmitry-kan.medium.com/ Dmitry on LinkedIn: https://www.linkedin.com/in/dmitrykan/
S4 · E4
15 giu 2026
56:51
AI Webinar - Building an AI-Ready Data Backbone
Webinar I gave with AI Camp and Aiven on AI-ready data backbone, and specifically how OpenSearch unlocks AI-powered search and log analytics: https://www.aicamp.ai/event/eventdetails/W2026032610 Blog post: https://dmitry-kan.medium.com/webinar-building-an-ai-ready-data-backbone-with-aiven-google-cloud-4629f97f69bd LLM/RAG/AI Agents course: https://dmitry-kan.medium.com/course-large-language-models-and-generative-ai-for-nlp-2025-98e31780de30 Free tier OpenSearch: https://aiven.io/free-opensearch Time codes: 1:01 Dima's intro + Vector Podcast 4:56 About Aiven 7:06 Why best? - Question from the audience 10:22 Free Tier OpenSearch! 11:57 Aiven's unifed platform 12:58 OpenSearch: What and Why 17:00 Why OpenSearch is AI-Ready? 18:26 What Aiven's OpenSearch gives you 20:44 Lexical vs semantic search 22:51 Technical use cases of OpenSearch 24:17 Reference Architecture with Kafka as event processor, and OpenSearch as storage and search layer 25:37 Aiven's case studies for OpenSearch 26:27 When to choose OpenSearch? 28:21 Demo of OpenSearch query UI 32:12 Is there any advantage in using Qdrant over OpenSearch? - Question from the audience 34:30 What is the vector lenght (in this demo)? - Question from the audience 36:27 What are the main advantages of Aiven's OpenSearch compared to Elasticsearch? - Question from the audience 32:11 Demo of Search Relevancy Workbench: visual way of searching Show notes: - User Behaviour Insights: https://www.ubisearch.dev/ - Webinar's demo code part 1: Episode download / transcribe / index: https://github.com/dimakan-dev/conduit-transcripts/blob/main/DATA_PROCESSING_GUIDE.md - Webinar's demo code part 2: Main UI and quality dashboards: https://github.com/dimakan-dev/preparing-data-for-opensearch-and-rag/blob/main/workshop/STREAMLIT_README.md
S4 · E3
10 apr 2026
1:18:25
Trey Grainger - Wormhole Vectors
This lightning session introduces a new idea in vector search - Wormhole vectors! It has deep roots in physics and allows for transcending spaces of any nature: sparse, vector and behaviour (but could theoretically be any N-dimensional space). Craft decaf & half caf coffee, 25% discount: https://savorista.com/discount/VECTOR Blog post on Medium: https://dmitry-kan.medium.com/novel-idea-in-vector-search-wormhole-vectors-6093910593b8 Session page on maven: https://maven.com/p/8c7de9/beyond-hybrid-search-with-wormhole-vectors?utm_campaign=NzI2NzIx&utm_medium=ll_share_link&utm_source=instructor To try the managed OpenSearch (multi-cloud, automatic backups, disaster recovery, vector search and more), go here: https://console.aiven.io/signup?utm_source=youtube&utm_medium&&utm_content=vectorpodcast Get credits to use Aiven's products (PG, Kafka, Valkey, OpenSearch, ClickHouse): https://aiven.io/startups Timecodes: 00:00 Intro by Dmitry 01:48 Trey's presentation 03:05 Walk to the AI-Powered Search course by Trey and Doug 07:07 Intro to vector spaces and embeddings 19:03 Disjoint vector spaces and the need of hybrid search 23:11 Different modes of search 24:49 Wormhole vectors 47:49 Q&A What you'll learn: - What are "Wormhole Vectors"? Learn how wormhole vectors work & how to use them to traverse between disparate vector spaces for better hybrid search. - Building a behavioral vector space from click stream data Learn to generate behavioral embeddings to be integrated with dense/semantic and sparse/lexical vector queries. - Traverse lexical, semantic, & behavioral vectors spaces Jump back and forth between multiple dense and sparse vector spaces in the same query - Advanced hybrid search techniques (beyond fusion algorithms) Hybrid search is more than mixing lexical + semantic search. See advanced techniques and where wormhole vectors fit in. YouTube: https://www.youtube.com/watch?v=fvDC7nK-_C0
S4 · E2
7 nov 2025
1:19:17
Economical way of serving vector search workloads with Simon Eskildsen, CEO Turbopuffer
Turbopuffer search engine supports such products as Cursor, Notion, Linear, Superhuman and Readwise. Craft decaf & half caf coffee, 25% discount: https://savorista.com/discount/VECTOR This episode on YouTube: https://youtu.be/I8Ztqajighg Medium: https://dmitry-kan.medium.com/vector-podcast-simon-eskildsen-turbopuffer-69e456da8df3 Dev: https://dev.to/vectorpodcast/vector-podcast-simon-eskildsen-turbopuffer-cfa If you are on Lucene / OpenSearch stack, you can go managed by signing up here: https://console.aiven.io/signup?utm_source=youtube&utm_medium=&&utm_content=vectorpodcast Time codes: 00:00 Intro 00:15 Napkin Problem 4: Throughput of Redis 01:35 Episode intro 02:45 Simon's background, including implementation of Turbopuffer 09:23 How Cursor became an early client 11:25 How to test pre-launch 14:38 Why a new vector DB deserves to exist? 20:39 Latency aspect 26:27 Implementation language for Turbopuffer 28:11 Impact of LLM coding tools on programmer craft 30:02 Engineer 2 CEO transition 35:10 Architecture of Turbopuffer 43:25 Disk vs S3 latency, NVMe disks, DRAM 48:27 Multitenancy 50:29 Recall@N benchmarking 59:38 filtered ANN and Big-ANN Benchmarks 1:00:54 What users care about more (than Recall@N benchmarking) 1:01:28 Spicy question about benchmarking in competition 1:06:01 Interesting challenges ahead to tackle 1:10:13 Simon's announcement Show notes: - Turbopuffer in Cursor: https://www.youtube.com/watch?v=oFfVt3S51T4&t=5223s transcript: https://lexfridman.com/cursor-team-transcript - https://turbopuffer.com/ - Napkin Math: https://sirupsen.com/napkin - Follow Simon on X: https://x.com/Sirupsen - Not All Vector Databases Are Made Equal: https://towardsdatascience.com/milvus-pinecone-vespa-weaviate-vald-gsi-what-unites-these-buzz-words-and-what-makes-each-9c65a3bd0696/
S4 · E1
19 set 2025
1:15:26
Stagione 3
Adding ML layer to Search: Hybrid Search Optimizer with Daniel Wrigley and Eric Pugh
Vector Podcast website: https://vectorpodcast.com Haystack US 2025: https://haystackconf.com/2025/ Federated search, Keyword & Neural Search, ML Optimisation, Pros and Cons of Hybrid search It is fascinating and funny how things develop, but also turn around. In 2022-23 everyone was buzzing about hybrid search. In 2024 the conversation shifted to RAG, RAG, RAG. And now we are in 2025 and back to hybrid search - on a different level: finally there are strides and contributions towards making hybrid search parameters learnt with ML. How cool is that? Design: Saurabh Rai, https://www.linkedin.com/in/srbhr/ The design of this episode is inspired by a scene in Blade Runner 2049. There's a clear path leading towards where people want to go to, yet they're searching for something. 00:00 Intro 00:54 Eric's intro and Daniel's background 02:50 Importance of Hybrid search: Daniel's take 07:26 Eric's take 10:57 Dmitry's take 11:41 Eric's predictions 13:47 Doug's blog on RRF is not enough 16:18 How to not fall short of the blind picking in RRF: score normalization, combinations and weights 25:03 The role of query understanding: feature groups 35:11 Lesson 1 from Daniel: Simple models might be all you need 36:30 Lesson 2: query features might be all you need 38:30 Reasoning capabilities in search 40:02 Question from Eric: how is this different from Learning To Rank? 42:46 Carrying the past in Learning To Rank / any rank 44:21 Demo! 51:52 How to consume this in OpenSearch 55:15 What's next 58:44 Haystack US 2025 YouTube: https://www.youtube.com/watch?v=quY769om1EY
S3 · E11
21 mar 2025
1:03:09
Vector Databases: The Rise, Fall and Future - by NotebookLM
https://www.vectorpodcast.com/ I had fun interacting with NotebookLM - mostly for self-educational purposes. I think this tool can help by bringing an additional perspective over a textual content. It ties to what RAG (Retrieval Augmented Generation) can do to content generation in another modality. In this case, text is used to augment the generation of a podcast episode. This episode is based on my blog post: https://dmitry-kan.medium.com/the-rise-fall-and-future-of-vector-databases-how-to-pick-the-one-that-lasts-6b9fbb43bbbe Time codes: 00:00 Intro to the topic 1:11 Dmitry's knowledge in the space 1:54 Unpacking the Rise & Fall idea 3:14 How attention got back to Vector DBs for a bit 4:18 Getting practical: Dmitry's guide for choosing the right Vector Database 4:39 FAISS 5:34 What if you need fine-grained keyword search? Look at Apache Lucene-based engines 6:41 Exception to the rule: Late-interaction models 8:30 Latency and QPS: GSI APU, Vespa, Hyperspace 9:28 Strategic approach 9:55 Cloud solutions: CosmosDB, Vertex AI, Pinecone, Weaviate Cloud 10:14 Community voice: pgvector 10:48 Picture of the fascinating future of the field 12:23 Question to the audience 12:44 Taking a step back: key points 13:45 Don't get caught up in trendy shiny new tech YouTube: https://www.youtube.com/watch?v=403rxbWZK9Y
S3 · E10
2 mar 2025
19:52
Code search, Copilot, LLM prompting with empathy and Artifacts with John Berryman
Vector Podcast website: https://vectorpodcast.com Get your copy of John's new book "Prompt Engineering for LLMs: The Art and Science of Building Large Language Model–Based Applications": https://amzn.to/4fMj2Ef John Berryman is the founder and principal consultant of Arcturus Labs, where he specializes in AI application development (Agency and RAG). As an early engineer on GitHub Copilot, John contributed to the development of its completions and chat functionalities, working at the forefront of AI-assisted coding tools. John is coauthor of "Prompt Engineering for LLMs" (O'Reilly).Before his work on Copilot, John's focus was search technology. His diverse experience includes helping to develop next-generation search system for the US Patent Office, building search and recommendations for Eventbrite, and contributing to GitHub's code search infrastructure. John is also coauthor of "Relevant Search" (Manning), a book that distills his expertise in the field.John's unique background, spanning both cutting-edge AI applications and foundational search technologies, positions him at the forefront of innovation in LLM applications and information retrieval. 00:00 Intro 02:19 John's background and story in search and ML 06:03 Is RAG just a prompt engineering technique? 10:15 John's progression from a search engineer to ML researcher 13:40 LLM predictability vs more traditional programming 22:31 Code assist with GitHub Copilot 29:44 Role of keyword search for code at GitHub 35:01 GenAI: existential risk or pure magic? AI Natives 39:40 What are Artifacts 46:59 Demo! 55:13 Typed artifacts, tools, accordion artifacts 56:21 From Web 2.0 to Idea exchange 57:51 Spam will transform into Slop 58:56 John's new book and Acturus Labs intro Show notes: - John Berryman on X: https://x.com/JnBrymn - Acturus Labs: https://arcturus-labs.com/ - John's blog on Artifacts (see demo in the episode): https://arcturus-labs.com/blog/2024/11/11/cut-the-chit-chat-with-artifacts/ YouTube: https://youtu.be/60HAtHVBYj8
S3 · E9
10 feb 2025
1:07:24
Debunking myths of vector search and LLMs with Leo Boytsov
00:00 Intro 01:31 Leo's story 09:59 SPLADE: single model to solve both dense and sparse? 21:06 DeepImpact 29:58 NMSLIB: what are non-metric spaces 34:21 How HNSW and NMSLIB joined forces 41:11 Why FAISS did not choose NMSLIB's algorithm 43:36 Serendipity of discovery and the creation of industries 47:06 Vector Search: intellectually rewarding, professionally undervalued 52:37 Why RDBMS Still Struggles with Scalable Vector and Free-Text Search 1:00:16 Leo's recent favorite papers Leo Boytsov on LinkedIn: https://www.linkedin.com/in/leonidboytsov/ and X: https://x.com/srchvrs Leo Boytsov’s paper list: https://scholar.google.com/citations?hl=en&user=I79y2i4AAAAJ&view_op=list_works&sortby=pubdate Lots of papers and other material from Leo: https://www.youtube.com/watch?v=gzWErcOXIKk
S3 · E8
17 gen 2025
1:07:54
Berlin Buzzwords 2024 - Alessandro Benedetti - LLMs in Solr
This episode on YouTube: https://www.youtube.com/watch?v=PNB70TbQUBE Alessandro's talk on Hybrid Search with Apache Solr Reciprocal Rank Fusion: https://www.youtube.com/watch?v=8x2cbT5CCEM&list=PLq-odUc2x7i8jHpa6PHGzmxfAPEz-c-on&index=5 00:00 Intro 00:50 Alessandro's take on the bbuzz'24 conference 01:25 What and value of hybrid search 04:55 Explainability of vector search results to users 09:27 Explainability of vector search results to search engineers 13:12 State of hybrid search in Apache Solr 14:32 What's in Reciprocal Rank Fusion beyond round-robin? 18:30 Open source for LLMs 22:48 How we should approach this issue in business and research 26:12 How to maintain the status of an open-source LLM / system 30:06 Prompt engineering (hope and determinism) 34:03 DSpy 35:16 What's next in Solr
S3 · E7
7 nov 2024
38:04
1 di 4