ms-marco

AuraVector DB: High-precision, low-latency hybrid vector engine for RAG systems. Combines BGE dense embeddings, Rank-BM25, Reciprocal Rank Fusion (RRF), Cross-Encoder reranking, and SVD entropy analysis to eliminate LLM hallucinations under a p95 < 300 ms constraint. Built for ADROSONIC Build Challenge by Team SynaptiX (BIT Mesra).

Diploma thesis evaluating Retrieval-Augmented Generation (RAG) systems on MS MARCO and LegalBench-RAG using RAGChecker and RAGAs.

Python client for the ltr-bert-sir ELSIE API (sir-elsie Space)

TypeScript client for the ltr-bert-sir ELSIE API (sir-elsie Space)

Sign-coded late-interaction retrieval at 8 B/token: code, paper and full reproduction recipe for ReNeuIR @ SIGIR 2026

FlashRank-Pro: 149M-parameter ModernBERT cross-encoder reranker for RAG. Apache 2.0. Measured 0.3314 five-set BEIR-style NDCG@10 (+22.6% vs run-init). CPU-friendly, 8K context.

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).

Concept Model g(x)=W·SPLADE(x) for sparse-to-dense retrieval on MS MARCO and BEIR

Two-stage neural search ranking on MS MARCO: BM25 candidate generation + a fine-tuned Transformer cross-encoder re-ranker. +31% MRR@10 over BM25 on held-out queries, with a re-ranking depth vs. quality analysis.

Fine-tuning cross-encoder re-rankers on MS MARCO, fusing their rankings, and expanding queries with an LLM, evaluated on TREC DL'19

Local ONNX cross-encoder reranking for JavaScript/TypeScript. Zero API costs, framework-agnostic, with a Vercel-style API.

Evaluating the effect of hard and soft watermarking on LLM text quality using centroid-based embedding analysis across 15,000 samples from the MS MARCO dataset. APPS597 Supervised Independent Study - University of Otago.

Small-scale replication of ColBERT late-interaction retrieval with BERT, MaxSim, MS MARCO training, indexing, and retrieval evaluation.

Multi-teacher LLM distillation for document reranking - three open-source teachers (Qwen2.5-14B-Instruct, DeepSeek-R1-Distill-Llama-8B, Mistral-Nemo-12B) teach a 7B student to rank and explain its rankings via QLoRA. Evaluated on TREC DL19/DL20.

Multi-stage retrieval on MS MARCO (8.8M passages): BM25 + MonoT5 re-ranking + T5 query rewriting. NDCG@3 improved from 0.098 to 0.274.

A compact research project on LLM-based query understanding for search: query rewriting, intent and entity extraction, robustness to noisy queries, and downstream retrieval evaluation.