Advanced RAG 01: Small-to-Big Retrieval with LlamaIndex

Two primary techniques:
- Smaller Child Chunks Referring to Bigger Parent Chunks: Fetch smaller chunks during retrieval first, then reference the parent IDs, and return the bigger chunks.
- Sentence Window Retrieval: Fetch a single sentence during retrieval and return a window of text around the sentence.
🔗 Blog: / advanced-rag-01-small-...
🔗 Notebook: colab.research.google.com/git...
🔔 SUBSCRIBE to my channel: kzbin.info...
⭐ Stay in touch ⭐
📚 DS/ML Book Club: dsbookclub.github.io/
▶ KZbin: / sophiayangds
✍️ Medium: / sophiamyang
🐦 Twitter: / sophiamyang
🤝 Linkedin: / sophiamyang
💚 #ai

[@danieltost]

Loved the small to big retrieval. Simple and elegant. Awesome work Sophia!!

[@sivi3883]

Thanks for sharing! Cannot wait to try this and see the difference in the quality of RAG output. Excited to hear more tricks in this series!!!

[@assouliyoussef9912]

just brilliant , thank you

[@MakeStuffWithAI]

Awesome tutorial!

[@whlau6191]

Sophia, thanks for your sharing! Your presentation is always clear, structural and target oriented. it is so nice that you are one of the speakers in HK pycon. See you! 😊

[@JamesRBentley]

Another great video👍

[@olliegooding]

great explanations!

[@chukypedro818]

Awesome, I've tried your approach to RAG, and it is far better than the current Langchain RAG that I am using. You are the best, Sophia.😘

[@TheCherryo]

I thought of something similar to this but never had the time to implement it myself. I am very excited to see all the advancements in RAG and Agent-like applications. Curious to see if there is going to be specialized models that work better with RAG applications.

[@klaudioz_]

Amazing content!!. Thanks for sharing 🚀

[@SophiaYangDS]

Thank you 🙏😊

[@ztq121121]

Thank you for sharing this! I'm curious if there are any benchmarks or evaluations that quantify the extent of improvement that small-to-big retrieval can offer. Is there any research paper that we can use as reference?

[@AkarshanBiswas]

Actually I do not use llama_index. I have created a separate class which does cosine similarity search for my local LLM's internet search feature.. This small chunks big chunks method looks very interesting. I will watch this video again tomorrow. Very good information!! Thank you.

[@SophiaYangDS]

Nice! Thanks! Yeah the small to big retrieval methods are useful 🙌

[@eklok5000]

Really interesting content! I am unfortunately not all that familiar with llms and those techniques so it is sometimes hard to follow. Do you have some resources to catch up on the fundamentals?

[@crowofhermes]

Hi first of all, thank you for the great explanations! may I ask how does the first recursive retriever differs from utilizing `HierarchicalNodeParser + AutoMergingRetriever`? I will check in detail to see if there is something else but they seem pretty similar

[@dhavalthakkar5147]

Sophia, is there any langchain implementation for the above techniques?

[@amitk3098]

Is there anyway I can implement this without langchain and llama-index? I am not in favor having my RAG solution dependent on them.

[@avotchiezoa6627]

where can i find your background video ? Which apps do you use for power point?

[@user-ow4kc4fo4c]

How to compare with graph llama index?

[@AbolfazlMahmoodi]

How to reduce cost of openai API call?

[@VenkatesanVenkat-fd4hg]

Superr video, can you discuss multimodal retriever system where docs or pdf contains tables & text.

[@SophiaYangDS]

Yes thanks for the suggestions. These are on my list 🙌

[@liju7024]

I still don't understand the connection between child chunck and parent chunck. Can you explain more about it?

[@sjsadjf7030]

Amazing video but I have a question. I would be very happy if you could respond; even though it enhances the accuracy a lot, recursive retriever takes so much time. Would persisting indexes to the disk also be a solution in this case? or is there an alternative way to reduce the time? Thank you very much!!

[@sjsadjf7030]

oh, when I switched the embed model to default (ada-002-v2) it is now much faster, now I understand that it is not recursion related lol :D the main question has changed to how to reduce the embedding model costs haha

[@Aldraz]

There are times when some new AI technique comes up and I am like yay.. but then I realize it's something I had implemented a year ago xD

[@user-wr4yl7tx3w]

Is this idea associated with a publication?

[@oguzhanzobar1094]

looks pretty cool, but I keep getting a ValueError: "Got a larger chunk overlap (200) than chunk size (128), should be smaller." when I try to run the code under "Chunk References: Smaller Child Chunks Referring to Bigger Parent Chunk".

[@nathanredin7785]

add ",chunk_overlap=20" in your splitter