-
The Slow Half of PLAID: Making Exact Rescoring Cheap with One Int8 Operation
Exact rescoring has always been the slow half of late-interaction search. One asymmetric int8 operation and a stage-1 rework made next-plaid 1.7.0 4.3 to 5.4x faster on CPU at matched quality, on existing indexes.
-
My Attempt at Building a Multi-GPU Embedding Server with TEI and Qdrant
Why there’s no one-click multi-GPU embedding server in 2025, and the TEI + Nginx + FastAPI + Qdrant stack I ended up building instead.
-
Reproducibility Report of ModernBERT Models for Retrieval Tasks Using DPR
Fine-tuning ModernBERT-base and -large on MS MARCO with DPR. The reproduced NDCG@10 beats the paper’s reported numbers on nearly every dataset, likely due to batch size.
Blog
Notes on retrieval, embeddings, and the infrastructure around them. Mostly things I hit while building, written down while they were still fresh.