Shubrashankh Chatterjee’s Post

Retrieved chunk correction is an interesting idea to try out within a RAG pipeline.

I am working on the final review: Which version is better, and how can you create your own rules and uncover hidden secrets with output results comparison.. 1st blog link here : https://lnkd.in/gPtkMkkr

A cool trick for improving retrieval quality for RAG is to include retrieval-evals in-the-loop. Given a set of retrieved results, use an LLM evaluator to decide how relevant each context is to the query, before synthesizing an answer. You can use this to filter results, augment context from “backup” sources, and more. This core idea was proposed in the recent CRAG paper (Corrective Retrieval Augmented Generation) by Yan et al., and now it’s available as a LlamaPack thanks to Ravi Theja Desetty! Check it out 👇 https://lnkd.in/gwj6_zDW Paper: https://lnkd.in/gtbwiE2C

Interesting article about further improving context retrieval for RAG applications. It introduces additional steps that distill and disambiguate retrieved data through reasoning and additional search. Worth a read! HeyIris.AI

A inspector or reviewer is always welcome who can tell what can be improved or what is not correct here it's goes CRAG to facilitate it

following this paper for the theory and implementing the code accordingly : https://lnkd.in/gDYyjB8K

Excited to share that our paper titled "Towards Reliable Latent Knowledge Estimation in LLMs: Zero-Prompt Many-Shot Based Factual Knowledge Extraction" has been accepted at WSDM 2025! This work has been led by Qinyuan Wu and co-authored by Mohammad Aflah Khan, Soumi Das, Vedant Nanda , Bishwamittra Ghosh, Camila Kolling, Till Speicher, Laurent Bindschaedler, Krishna Gummadi and Evimaria Terzi! The paper focuses on the challenging task of reliably estimating factual knowledge that is embedded inside LLMs. We propose to eliminate prompt engineering when probing LLMs for factual knowledge and term our approach as Zero-Prompt Latent Knowledge Estimator (ZP-LKE). This leverages the in-context learning ability of LLMs to communicate both the factual knowledge question as well as the expected answer format. We evaluated our method on 49 open-source LLMs from various families like Llama(2), Gemma, Mistral, OPT, and Pythia and found it to outperform previous prompt-based approaches. More details and preprint to follow soon.

Come learn all about tree-structured retrieval 🌲 RAPTOR is a recent paper that introduces tree-structured indexing/retrieval, which hierarchically clusters/summarizes chunks into a tree structure containing both high-level and low-level pieces. Compared to naive top-k RAG, this allows you to retrieve both low-level and high-level details to answer different questions. We're excited to host the authors: Parth Sarthi, Salman Abdullah, Shubh Khanna et al. for a special webinar. The authors will be presenting the paper itself and we'll do a Q&A talking about tree-structured retrieval more generally. This Thursday 9am PT: https://lu.ma/9vzrl7m5 We also have a native LlamaIndex implementation, check it out 👇 Pack: https://lnkd.in/gUn-QY-j Notebook: https://lnkd.in/gcu9prRT Source paper: https://lnkd.in/gHrZNyz7

Why does time appear to flow in one direction, a.k.a. Arrow of Time? Explained in recent paper (it’s free) https://lnkd.in/gd6m__db via simple examples. 

✒️ RECALL this paper📄 use NEAT to build a population of networks optimized for mulitclass classification problems you can review NEAT paper summary https://lnkd.in/daEVeep2 Why NEAT was advantage ? ✅ The ability to evolve the structure of the network remains a major benefit of evolutionary approaches over optimization techniques such as backpropagation ✨ In this paper, NEAT is evaluated in several multiclass classification problems, and then extended via two ensemble approaches: One-vs-All and One-vs-One ✨These approaches decompose multiclass classification problems into a set of binary classification problems ✨ in which each binary problem is solved by an instance of NEAT. ✨These ensemble models exhibit reduced variance and increasingly superior accuracy as the number of classes increases. 🔎 Read more in this summary of this interseting paper 📄