Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Search Evolution - Keeping up with the hype?

Search Evolution - Keeping up with the hype?

This talk is an introduction into search and focuses on the recent search trends in the last years. Search is not a solved problem. We start with the basics like relevance, extending over to learning to rank and vector search and will - of course - also cover LLMs and what might come in the future.

Alexander Reelsen

May 25, 2023
Tweet

More Decks by Alexander Reelsen

Other Decks in Technology

Transcript

  1. Search Search Evolution
    The next generation of search engines...
    Alexander Reelsen
    [email protected] | @spinscale

    View Slide

  2. Learn about the trends in search engines
    Understand that this is a highly volatile market in the coming years
    Today's goal

    View Slide

  3. Status quo

    View Slide

  4. Speed (Search & Suggests)
    Scale (all the internet)
    Relevance
    Intent
    Personalization
    The power of search

    View Slide

  5. Text search
    Enterprise search
    Ecommerce search
    Log search
    Analytics
    Dashboards
    NLP
    Generative/Conversational Search
    Evolution of Use-Cases

    View Slide

  6. SQL: Does row r match query q ?
    Answer: /
    How well matches query q document d ?
    Answer: [0..∞]
    Scoring based on formula: TF/IDF , BM25
    Dependent on corpus
    Relevancy

    View Slide

  7. Recency
    Rating
    Popularity
    Past (searches/purchases)
    Individualization
    Ranking

    View Slide

  8. Trends

    View Slide

  9. SaaS
    Splitting storage and compute
    Using blob storage, segment replication
    Massive cost savings
    Going cloud native

    View Slide

  10. Scoring/Relevancy based on machine learning model
    Common: Reranking after first filtering
    Machine Learning models trained independently
    Learning to rank

    View Slide

  11. Vector search engines: translates content into vectors
    QDrant, Milvus, Weaviate, Pinecone, Deeplake, nucliadb
    Best model wins...
    Going hybrid: Will search engines add vector support or vector engines
    add search support?
    Vector Search

    View Slide

  12. SQLite: vector extension, FTS3/4 extension
    Postgres: PostgresML - full model management and querying in Postgres!
    Don't sleep on SQL engines!

    View Slide

  13. Distributed search across regions
    Search on your browser
    Search on your phone
    Check out OramaSearch
    Search on the edge

    View Slide

  14. ChatGPT

    View Slide

  15. blue dress with white stripes that has been shown on the last fashion
    week in milan
    summarize the quarterly earnings call, focus on numbers that differ
    strongly from the last three quarters
    Convert the following CDK snippet from Java to python
    Generative/Conversational search

    View Slide

  16. blue dress with white stripes requires image extraction
    last fashion week in milan requires external knowledge
    Your own dataset is not enough for a good search!
    Generative search - context

    View Slide

  17. View Slide

  18. View Slide

  19. View Slide

  20. View Slide

  21. Prompt to any

    View Slide

  22. futuristic skyline in neon
    colors with a futuristic
    looking tesla model 3 in the
    foreground
    Stable diffusion

    View Slide

  23. View Slide

  24. Large size, trained on massive datasets
    Open Source: Langchain
    Prompt engineering
    Classification, Question Answering, Summarization, Fill-mask, Translation
    Hallucination & Model bias
    Conversational memory
    Learning from queries (dangerous?)
    Agents for LLMs (execute a calculator, SQL query, use mechanical turk)
    LLMs

    View Slide

  25. Cars
    Mobile
    Voice based search

    View Slide

  26. Summary

    View Slide

  27. Search becomes hybrid: Will the existing search engines adapt?
    Search customization is expensive - A brief history of code search at
    GitHub
    Search engine becomes the commodity
    Rent your industry specific LLM!
    Privacy LLMs might be a thing
    Expect a lot of movement, lots of "AI integrations" and even more hot
    air...
    Summary

    View Slide

  28. View Slide

  29. Thank you!
    Q & A
    Alexander Reelsen
    [email protected] | @spinscale

    View Slide