Don't naive RAG do hybrid search instead (Pinecone Weaviate or pgvector + full text search & rerank)

Поделиться
HTML-код
  • Опубликовано: 25 дек 2024

Комментарии •

  • @devlearnllm
    @devlearnllm  5 месяцев назад +4

    Hey yall, in case you didn't get good full text search results like me, the CEO of Supabase (Paul Copplestone) sent me this to use instead: supabase.com/docs/guides/database/extensions/pgroonga

  • @pabloarroyo7952
    @pabloarroyo7952 3 месяца назад +2

    Watching this 2 months later. Great video, thanks for sharing

  • @alienPear
    @alienPear 3 месяца назад +1

    Thanks for sharing, bro! Greetings from Colombia

  • @JamesRBentley
    @JamesRBentley 5 месяцев назад +1

    Nice video sir. I have already been experimenting with the colab - sincerest thanks

  • @ironbondar
    @ironbondar 5 месяцев назад +1

    very good workshop. straight to the point

  • @gregmeldrum
    @gregmeldrum 5 месяцев назад +1

    Very informative! A great resource. Thanks for sharing your wealth of knowledge!!

  • @magnusjensen5867
    @magnusjensen5867 2 месяца назад +1

    Nice workshop, thank you for sharing! You mentioned early on that you tried decomposing your queries if they were multi-hop queries / abstract queries. Would you still suggest that approach or is there any new research specifically on this matter? Imagine a query in which a user want to retrieve information from multiple documents at get a comparison or summarization.

    • @devlearnllm
      @devlearnllm  2 месяца назад +2

      I'm still doing the same for my app, and what I'm hoping to do eventually is to prompt the query expansion step so it's expanding in a coherent way. E.g question is about how X affects Y -> find X, find Y

    • @magnusjensen5867
      @magnusjensen5867 2 месяца назад

      @@devlearnllm Thank you for your response. How exactly would you go about this? Have you played with knowledge graph (GraphRAG) like Neo4j etc?

  • @Phoenix-gi3gu
    @Phoenix-gi3gu 4 месяца назад +1

    For experimenting I would recommend using no database at all. You can simply use the cosine similarity (i.e. from torch functional) or quickly implement it and you are nearly done. Just use some argsort to get the best matches. It's like five lines of code or so. For easy store/load you can use pickle to serialize/unserialize the object that holds the embeddings. It is fast on CPU too, but of course you can run it on GPU without any bigger changes.
    No services required.

  • @ofrylivney367
    @ofrylivney367 5 месяцев назад +2

    Nice workshop! I'll definitely try out the hybrid search. Do you recon it'll work with nomic text embeddings and ollama?

  • @oamarkanji3153
    @oamarkanji3153 3 месяца назад

    Incredible content. Thank you.

  • @SandeeshCroos
    @SandeeshCroos Месяц назад

    Hey, great content! Thanks for sharing your knowledge. However, instead of just using tsvector in PostgreSQL, you can leverage sparse vector search by utilizing the pg_search extension, right?

    • @devlearnllm
      @devlearnllm  Месяц назад +1

      yup, they're both full text search. Or use pgroonga

  • @ThoughtfullySo
    @ThoughtfullySo 5 месяцев назад +1

    You should've tried Qdrant.

  • @ajkdrag
    @ajkdrag 4 месяца назад +1

    Hi. I have a video request. Is there a way to contact you?

    • @devlearnllm
      @devlearnllm  3 месяца назад

      tally.so/r/n9djRQ

    • @ajkdrag
      @ajkdrag 3 месяца назад

      @@devlearnllm done. Thanks

  • @vijishmadhavan6093
    @vijishmadhavan6093 4 месяца назад

    what happens if we use all the 25000 cases, will it work?

    • @devlearnllm
      @devlearnllm  4 месяца назад

      Most likely. Pinecone, Weaviate and pgvector are very performant.

  • @ArunKumar-bp5lo
    @ArunKumar-bp5lo 5 месяцев назад +1

    great

  • @artur50
    @artur50 5 месяцев назад

    is it possible to run it with Ollama?

  • @zuowang5185
    @zuowang5185 5 месяцев назад

    Is Openai embedding v3 model better than Bert?

    • @devlearnllm
      @devlearnllm  5 месяцев назад

      Hard to tell unless experiments are run.
      huggingface.co/spaces/mteb/leaderboard

  • @flor.7797
    @flor.7797 5 месяцев назад

    I just use Google 🙃