RAPTOR: Dynamic Tree-Structured Summaries with LangChain - Advanced RAG

Follow up video of this: • RAPTOR - Advanced RAG ...
Code: github.com/Coding-Crashkurse/...
Paper: arxiv.org/html/2401.18059v1
Timestamps
0:00 What is RAPTOR?
1:02 Code Walkthrough
10:08 Dynamic Code in Action

[@RiskSeeking-gi4mi]

Thanks for this! I’ve been extending this to find values for n_neighbors, dims, etc that maximize the quality of the clusters. I’m applying this to 10-k filings which are pretty similar in overall “semantic” content and organization, so I’m hoping that as I process more 10-Ks, I’ll gradually find a set of parameters that generalizes well across most 10-ks. Kind of surprising that I haven’t seen anyone talk about this as far as RAPTOR

[@codingcrashcourses8533]

Have you had the chance to evaluate the performance of that compared to other retrieval techniques? would be interested in the results :)

[@patriciachirwa9183]

Hi, is line 157 in the code meant to come before the iteration summaries loop or after in line 165? i.e. are we updating the all_summaries field with the previous cluster texts or does it not matter? Otherwise we would be updating "iteration summaries["texts"]" with the same value as "iteration summaries["summaries"]"

[@micbab-vg2mu]

Great:)

[@henkhbit5748]

Thanks, why not using open source LLM and embeddings?

[@codingcrashcourses8533]

The state of the art models change so fast which is why I prefer OpenAI. But the code should stay pretty much the same and not really matter for this concept.

[@robertputneydrake]

Awesome. If I might suggest: how about a tutorial on CodeGen-specific advanced RAG, i.e. repository-wide code "understanding" and generation? :) Cheers!

[@codingcrashcourses8533]

I like the idea for a Video. But currently no Clue how i would tackle that yet

[@lesptitsoiseaux]

Is this subject to loss in the middle problem?

[@codingcrashcourses8533]

Yes, like any other ingestion step. You got methods like reranking to fight problems like this :)

[@loicbaconnier9150]

Ho it look like my idea in the previous video no ?
So you don’t need anymore the code ?

[@codingcrashcourses8533]

I did not create subclusters:)

[@loicbaconnier9150]

Whatever you do is always really interesting. Thanks for sharing