Query Expension for Better Query Embedding using LLMs

Code and models for the paper "Questions Are All You Need to Train a Dense Passage Retriever (TACL 2023)"

SPRINT Toolkit helps you evaluate diverse neural sparse models easily using a single click on any IR dataset.

Evaluation of BEIR Datasets using ColBERT retrieval model

A genral RAG Search chatbot, with SoTA RAG techniques such as HyDE, Hybrid retrieval with BM25 + RRF and Cross encoder reranking. Evaluated on the BEIR scifact dataset and compared all the different pipelines i tried along the way

Physics-Inspired Reranking via Token-Level Point Clouds & PDE Fusion | NFCorpus NDCG@10 = 0.3232 (+47.2%) | 26ms CPU | Zero training

The only way to accumulate, share, and stream knowledge seamlessly into your LLM work sessions. Peer-to-peer shared knowledge network — continuous shared learning across machines. Knowledge planes: everlastingly up-to-date, globally accessible, at lightning speed. libp2p, W3C did:key, local-first, MIT. 75.22% NDCG@10 on BEIR SciFact.

RAG evaluation on BEIR SciFact: BM25, dense and hybrid retrieval with LLM answers.

BM25 & SBERT retrieval on the FiQA-2018 financial QA benchmark · Gradio demo

Scripts to convert the LegalBench-RAG dataset into the standard IR format

Browser-side IR benchmark: BM25 vs Semantic vs Hybrid retrieval on SciFact (BEIR). Bauman MSTU NIR 2026.

A RAG system that replaces standard BM25/FAISS retrieval with a fully learned neural retrieval stack - including a fine-tuned bi-encoder, a cross-encoder reranker, ColBERT-style late interaction scoring, and a locally hosted LLM generator. Built entirely with free and open-source tools.

Controlled depth ablation of a BERT bi-encoder across training budgets and seeds on three BEIR tasks (nfcorpus, scifact, fiqa). L3–L12 is flat within seed noise at 20K steps; 80K training degrades every depth on zero-shot transfer (−45% NDCG@10 on fiqa for L12).

Given a set of documents and the minimum required similarity threshold find the number of document pairs that exceed the threshold

Retrieve the top-𝑘 documents with respect to a given query by maximal inner product over dense and sparse vectors

Comparative study of transformer and non-transformer encoder architectures for dense retrieval — mapping the latency × accuracy Pareto frontier on BEIR. Solo research, PES University CSE 2026.