Vector Podcast

Topics:

00:00 Intro

01:54 Things Connor learnt in the past year that changed his perception of Vector Search

02:42 Is search becoming conversational?

05:46 Connor asks Dmitry: How Large Language Models will change Search?

08:39 Vector Search Pyramid

09:53 Large models, data, Form vs Meaning and octopus underneath the ocean

13:25 Examples of getting help from ChatGPT and how it compares to web search today

18:32 Classical search engines with URLs for verification vs ChatGPT-style answers

20:15 Hybrid search: keywords + semantic retrieval

23:12 Connor asks Dmitry about his experience with sparse retrieval

28:08 SPLADE vectors

34:10 OOD-DiskANN: handling the out-of-distribution queries, and nuances of sparse vs dense indexing and search

39:54 Ways to debug a query case in dense retrieval (spoiler: it is a challenge!)

44:47 Intricacies of teaching ML models to understand your data and re-vectorization

49:23 Local IDF vs global IDF and how dense search can approach this issue

54:00 Realtime index

59:01 Natural language to SQL

1:04:47 Turning text into a causal DAG

1:10:41 Engineering and Research as two highly intelligent disciplines

1:18:34 Podcast search

1:25:24 Ref2Vec for recommender systems

1:29:48 Announcements

For Show Notes, please check out the YouTube episode below.

This episode on YouTube: https://www.youtube.com/watch?v=2Q-7taLZ374

Podcast design: Saurabh Rai: https://twitter.com/srvbhr

Toloka’s support for Academia: grants and educator partnerships

https://toloka.ai/collaboration-with-educators-form

https://toloka.ai/research-grants-form

These are pages leading to them:

https://toloka.ai/academy/education-partnerships

https://toloka.ai/grants

Topics:

00:00 Intro

01:25 Jenny’s path from graduating in ML to a Data Advocate role

07:50 What goes into the labeling process with Toloka

11:27 How to prepare data for labeling and design tasks

16:01 Jenny’s take on why Relevancy needs more data in addition to clicks in Search

18:23 Dmitry plays the Devil’s Advocate for a moment

22:41 Implicit signals vs user behavior and offline A/B testing

26:54 Dmitry goes back to advocating for good search practices

27:42 Flower search as a concrete example of labeling for relevancy

39:12 NDCG, ERR as ranking quality metrics

44:27 Cross-annotator agreement, perfect list for NDCG and Aggregations

47:17 On measuring and ensuring the quality of annotators with honeypots

54:48 Deep-dive into aggregations

59:55 Bias in data, SERP, labeling and A/B tests

1:16:10 Is unbiased data attainable?

1:23:20 Announcements

This episode on YouTube: https://youtu.be/Xsw9vPFqGf4

Podcast design: Saurabh Rai: https://twitter.com/srvbhr

00:00 Introduction

01:11 Yaniv’s background and intro to Searchium & GSI

04:12 Ways to consume the APU acceleration for vector search

05:39 Power consumption dimension in vector search

7:40 Place of the platform in terms of applications, use cases and developer experience

12:06 Advantages of APU Vector Search Plugins for Elasticsearch and OpenSearch compared to their own implementations

17:54 Everyone needs to save: the economic profile of the APU solution

20:51 Features and ANN algorithms in the solution

24:23 Consumers most interested in dedicated hardware for vector search vs SaaS

27:08 Vector Database or a relevance oriented application?

33:51 Where to go with vector search?

42:38 How Vector Search fits into Search

48:58 Role of the human in the AI loop

58:05 The missing bit in the AI/ML/Search space

1:06:37 Magical WHY question

1:09:54 Announcements

- Searchium vector search: https://searchium.ai/

- Dr. Avidan Akerib, founder behind the APU technology: https://www.linkedin.com/in/avidan-akerib-phd-bbb35b12/

- OpenSearch benchmark for performance tuning: https://betterprogramming.pub/tired-of-troubleshooting-idle-search-resources-use-opensearch-benchmark-for-performance-tuning-d4277c9f724

- APU KNN plugin for OpenSearch: https://towardsdatascience.com/bolster-opensearch-performance-with-5-simple-steps-ca7d21234f6b

- Multilingual and Multimodal Search with Hardware Acceleration: https://blog.muves.io/multilingual-and-multimodal-vector-search-with-hardware-acceleration-2091a825de78

- Muves talk at Berlin Buzzwords, where we have utilized GSI APU: https://blog.muves.io/muves-at-berlin-buzzwords-2022-3150eef01c4

- 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

Episode on YouTube: https://youtu.be/EerdWRPuqd4

Podcast design: Saurabh Rai: https://twitter.com/srvbhr

This episode on YouTube: https://www.youtube.com/watch?v=Kpua1Euc-B8

Topics:

00:00 Intro

01:30 Doug’s story in Search

04:55 How Quepid came about

10:57 Relevance as product at Shopify: challenge, process, tools, evaluation

15:36 Search abandonment in Ecommerce

21:30 Rigor in A/B testing

23:53 Turn user intent and content meaning into tokens, not words into tokens

32:11 Use case for vector search in Maps. What about search in other domains?

38:05 Expanding on dense approaches

40:52 Sparse, dense, hybrid anyone?

48:18 Role of HNSW, scalability and new vector databases vs Elasticsearch / Solr dense search

52:12 Doug’s advice to vector database makers

58:19 Learning to Rank: how to start, how to collect data with active learning, what are the ML methods and a mindset

1:12:10 Blending search and recommendation

1:16:08 Search engineer role and key ingredients of managing search projects today

1:20:34 What does a Product Manager do on a Search team?

1:26:50 The magical question of WHY

1:29:08 Doug’s announcements

Show notes:

Doug’s course: https://www.getsphere.com/ml-engineering/ml-powered-search?source=Instructor-Other-070922-vector-pod

Upcoming book: https://www.manning.com/books/ai-powered-search?aaid=1&abid=e47ada24&chan=aips

Doug’s post in Shopify’s blog “Search at Shopify—Range in Data and Engineering is the Future”: https://shopify.engineering/search-at-shopify

Doug’s own blog: https://softwaredoug.com/

Using Bayesian optimization for Elasticsearch relevance: https://www.youtube.com/watch?v=yDcYi-ANJwE&t=1s

Hello LTR: https://github.com/o19s/hello-ltr

Vector Databases: https://towardsdatascience.com/milvus-pinecone-vespa-weaviate-vald-gsi-what-unites-these-buzz-words-and-what-makes-each-9c65a3bd0696

Research: Search abandonment has a lasting impact on brand loyalty: https://cloud.google.com/blog/topics/retail/search-abandonment-impacts-retail-sales-brand-loyalty

Quepid: https://quepid.com/

Podcast design: Saurabh Rai [https://twitter.com/srvbhr]

YouTube: https://www.youtube.com/watch?v=N5Brb7Rzc2c

Topics:

00:00 Introduction

01:12 Malte’s background

07:58 NLP crossing paths with Search

11:20 Product discovery: early stage repetitive use cases pre-dating Haystack

16:25 Acyclic directed graph for modeling a complex search pipeline

18:22 Early integrations with Vector Databases

20:09 Aha!-use case in Haystack

23:23 Capabilities of Haystack today

30:11 Deepset Cloud: end-to-end deployment, experiment tracking, observability, evaluation, debugging and communicating with stakeholders

39:00 Examples of value for the end-users of Deepset Cloud

46:00 Success metrics

50:35 Where Haystack is taking us beyond MLOps for search experimentation

57:13 Haystack as a smart assistant to guide experiments

1:02:49 Multimodality

1:05:53 Future of the Vector Search / NLP field: large language models

1:15:13 Incorporating knowledge into Language Models & an Open NLP Meetup on this topic

1:16:25 The magical question of WHY

1:23:47 Announcements from Malte

Show notes:

- Haystack: https://github.com/deepset-ai/haystack/

- Deepset Cloud: https://www.deepset.ai/deepset-cloud

- Tutorial: Build Your First QA System: https://haystack.deepset.ai/tutorials/v0.5.0/first-qa-system

- Open NLP Meetup on Sep 29th (Nils Reimers talking about “Incorporating New Knowledge Into LMs”): https://www.meetup.com/open-nlp-meetup/events/287159377/

- Atlas Paper (Few shot learning with retrieval augmented large language models): https://arxiv.org/abs/2208.03299

- Zero click search: https://www.searchmetrics.com/glossary/zero-click-searches/

Very large LMs:

- 540B PaLM by Google: https://lnkd.in/eajsjCMr

- 11B Atlas by Meta: https://lnkd.in/eENzNkrG

- 20B AlexaTM by Amazon: https://lnkd.in/eyBaZDTy

- Players in Vector Search: https://www.youtube.com/watch?v=8IOpgmXf5r8 https://dmitry-kan.medium.com/players-in-vector-search-video-2fd390d00d6

- Click Residual: A Query Success Metric: https://observer.wunderwood.org/2022/08/08/click-residual-a-query-success-metric/

- Tutorials and papers around incorporating Knowledge into Language Models: https://cs.stanford.edu/people/cgzhu/

00:00 Introduction

01:10 Max's deep experience in search and how he transitioned from structured data

08:28 Query-term dependence problem and Max's perception of the Vector Search field

12:46 Is vector search a solution looking for a problem?

20:16 How to move embeddings computation from GPU to CPU and retain GPU latency?

27:51 Plug-in neural model into Java? Example with a Hugging Face model

33:02 Web-server Mighty and its philosophy

35:33 How Mighty compares to in-DB embedding layer, like Weavite or Vespa

39:40 The importance of fault-tolerance in search backends

43:31 Unit economics of Mighty

50:18 Mighty distribution and supported operating systems

54:57 The secret sauce behind Mighty's insane fast-ness

59:48 What a customer is paying for when buying Mighty

1:01:45 How will Max track the usage of Mighty: is it commercial or research use?

1:04:39 Role of Open Source Community to grow business

1:10:58 Max's vision for Mighty connectors to popular vector databases

1:18:09 What tooling is missing beyond Mighty in vector search pipelines

1:22:34 Fine-tuning models, metric learning and Max's call for partnerships

1:26:37 MLOps perspective of neural pipelines and Mighty's role in it

1:30:04 Mighty vs AWS Inferentia vs Hugging Face Infinity

1:35:50 What's left in ML for those who are not into Python

1:40:50 The philosophical (and magical) question of WHY

1:48:15 Announcements from Max

25% discount for the first year of using Mighty in your great product / project with promo code VECTOR:

https://bit.ly/3QekTWE

Show notes:

- Max's blog about BERT and search relevance: https://opensourceconnections.com/blog/2019/11/05/understanding-bert-and-search-relevance/

- Case study and unit economics of Mighty: https://max.io/blog/encoding-the-federal-register.html

- 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

Watch on YouTube: https://youtu.be/LnF4hbl1cE4

YouTube: https://www.youtube.com/watch?v=r4HEpyur-OE

Vector Podcast Live

Topics:

00:00 Kick-off introducing co:rise study platform

03:03 Grant’s background

04:58 Principle of 3 C’s in the life of a CTO: Code, Conferences and Customers

07:16 Principle of 3 C’s in the Search Engine development: Content, Collaboration and Context

11:51 Balance between manual tuning in pursuit to learn and Machine Learning

15:42 How to nurture intuition in building search engine algorithms

18:51 How to change the approach of organizations to true experimentation

23:17 Where should one start in approaching the data (like click logs) for developing a search engine

29:36 How to measure the success of your search engine

33:50 The role of manual query rating to improve search result relevancy

36:56 What are the available datasets, tools and algorithms, that allow us to build a search engine?

41:56 Vector search and its role in broad search engine development and how the profession is shaping up

49:01 The magical question of WHY: what motivates Grant to stay in the space

52:09 Announcement from Grant: course discount code DGSEARCH10

54:55 Questions from the audience

Show notes:

- Grant’s interview at Berlin Buzzwords 2016: https://www.youtube.com/watch?v=Y13gZM5EGdc

- “BM25 is so Yesterday: Modern Techniques for Better Search”: https://www.youtube.com/watch?v=CRZfc9lj7Po

- “Taming text” - book co-authored by Grant: https://www.manning.com/books/taming-text

- Search Fundamentals course - https://corise.com/course/search-fundamentals

- Search with ML course - https://corise.com/course/search-with-machine-learning

- Click Models for Web Search: https://github.com/markovi/PyClick

- Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing, book by Ron Kohavi et al: https://www.amazon.com/Trustworthy-Online-Controlled-Experiments-Practical-ebook/dp/B0845Y3DJV

- Quepid, open source tool and free service for query rating and relevancy tuning: https://quepid.com/

- Grant’s talk in 2013 where he discussed the need of a vector field in Lucene and Solr: https://www.youtube.com/watch?v=dCCqauwMWFE

- Demo of multimodal search with CLIP: https://blog.muves.io/multilingual-and-multimodal-vector-search-with-hardware-acceleration-2091a825de78

- Learning to Boost: https://www.youtube.com/watch?v=af1dyamySCs

YouTube: https://www.youtube.com/watch?v=GyXggc4LNKI

Topics:

00:00 Kick-off by Judy Zhu

01:33 Introduction by Dmitry Kan and his bio!

03:03 Daniel’s background

04:46 “Science is the difference between instinct and strategy”

07:41 Search as a personal learning experience

11:53 Why do we need Machine Learning in Search, or can we use manually curated features?

16:47 Swimming up-stream from relevancy: query / content understanding and where to start?

23:49 Rule-based vs Machine Learning approaches to Query Understanding: Pareto principle

29:05 How content understanding can significantly improve your search engine experience

32:02 Available datasets, tools and algorithms to train models for content understanding

38:20 Daniel’s take on the role of vector search in modern search engine design as the path to language of users

45:17 Mystical question of WHY: what drives Daniel in the search space today

49:50 Announcements from Daniel

51:15 Questions from the audience

Show notes:

[What is Content Understanding?. Content understanding is the foundation… | by Daniel Tunkelang | Content Understanding | Medium](https://medium.com/content-understanding/what-is-content-understanding-4da20e925974)

Query Understanding: An Introduction https://queryunderstanding.com/introduction-c98740502103)

Science as Strategy [YouTube](https://www.youtube.com/watch?v=dftt6Yqgnuw)

Search Fundamentals course - https://corise.com/course/search-fundamentals

Search with ML course - https://corise.com/course/search-with-machine-learning

Books:

Faceted Search, by Daniel Tunkelang: https://www.amazon.com/Synthesis-Lectures-Information-Concepts-Retrieval/dp/1598299999

Modern Information Retrieval: The Concepts and Technology Behind Search, by Ricardo Baeza-Yates: https://www.amazon.com/Modern-Information-Retrieval-Concepts-Technology/dp/0321416910/ref=sr11?qid=1653144684&refinements=p_27%3ARicardo+Baeza-Yates&s=books&sr=1-1

Introduction to Information Retrieval, by Chris Manning: https://www.amazon.com/Introduction-Information-Retrieval-Christopher-Manning/dp/0521865719/ref=sr1fkmr0_1?crid=2GIR19OTZ8QFJ&keywords=chris+manning+information+retrieval&qid=1653144967&s=books&sprefix=chris+manning+information+retrieval%2Cstripbooks-intl-ship%2C141&sr=1-1-fkmr0

Query Understanding for Search Engines, by Yi Chang and Hongbo Deng: https://www.amazon.com/Understanding-Search-Engines-Information-Retrieval/dp/3030583333

YouTube: https://www.youtube.com/watch?v=AU0O_6-EY6s

Topics:

00:00 Intro

01:03 Yusuf’s background

03:00 Multimodal search in tech and humans

08:53 CLIP: discovering hidden semantics

13:02 Where to start to apply metric learning in practice. AutoEncoder architecture included!

19:00 Unpacking it further: what is metric learning and the difference with deep metric learning?

28:50 How Deep Learning allowed us to transition from pixels to meaning in the images

32:05 Increasing efficiency: vector compression and quantization aspects

34:25 Yusuf gives a practical use-case with Conversational AI of where metric learning can prove to be useful. And tools!

40:59 A few words on how the podcast is made :) Yusuf’s explanation of how Gmail smart reply feature works internally

51:19 Metric learning helps us learn the best vector representation for the given task

52:16 Metric learning shines in data scarce regimes. Positive impact on the planet

58:30 Yusuf’s motivation to work in the space of vector search, Qdrant, deep learning and metric learning — the question of Why

1:05:02 Announcements from Yusuf

- Join discussions at Discord: https://discord.qdrant.tech

- Yusuf's Medium: https://medium.com/@yusufsarigoz and LinkedIn: https://www.linkedin.com/in/yusufsarigoz/

- GSOC 2022: TensorFlow Similarity - project led by Yusuf: https://docs.google.com/document/d/1fLDLwIhnwDUz3uUV8RyUZiOlmTN9Uzy5ZuvI8iDDFf8/edit#heading=h.zftd93u5hfnp

- Dmitry's Twitter: https://twitter.com/DmitryKan

Full Show Notes: https://www.youtube.com/watch?v=AU0O_6-EY6s

Topics:

00:00 Introduction

01:21 Jo Kristian’s background in Search / Recommendations since 2001 in Fast Search & Transfer (FAST)

03:16 Nice words about Trondheim

04:37 Role of NTNU in supplying search talent and having roots in FAST

05:33 History of Vespa from keyword search

09:00 Architecture of Vespa and programming language choice: C++ (content layer), Java (HTTP requests and search plugins) and Python (pyvespa)

13:45 How Python API enables evaluation of the latest ML models with Vespa and ONNX support

17:04 Tensor data structure in Vespa and its use cases

22:23 Multi-stage ranking pipeline use cases with Vespa

24:37 Optimizing your ranker for top 1. Bonus: cool search course mentioned!

30:18 Fascination of Query Understanding, ways to implement and its role in search UX

33:34 You need to have investment to get great results in search

35:30 Game-changing vector search in Vespa and impact of MS Marco Passage Ranking

38:44 User aspect of vector search algorithms

43:19 Approximate vs exact nearest neighbor search tradeoffs

47:58 Misconceptions in neural search

52:06 Ranking competitions, idea generation and BERT bi-encoder dream

56:19 Helping wider community through improving search over CORD-19 dataset

58:13 Multimodal search is where vector search shines

1:01:14 Power of building fully-fledged demos

1:04:47 How to combine vector search with sparse search: Reciprocal Rank Fusion

1:10:37 The philosophical WHY question: Jo Kristian’s drive in the search field

1:21:43 Announcement on the coming features from Vespa

- Jo Kristian’s Twitter: https://twitter.com/jobergum

- Dmitry’s Twitter: https://twitter.com/DmitryKan

For the Show Notes check: https://www.youtube.com/watch?v=UxEdoXtA9oM