The Single Integer That Was Quietly Wrecking Our Retrieval

Reciprocal Rank Fusion’s default k=60 comes from a 2009 web-search paper. On a small technical corpus it flattens your ranking to noise. Drop k to 15 and top results get rewarded again.
The Component We Almost Threw Away: Why a Reranker Is the Gatekeeper of Good RAG

A cross-encoder reranker moved our benchmark from 70% to 78%, but only in combination. Two industry-standard additions dropped accuracy 35 points on their own. What the reranker is actually for.
Everyone Argues About Which AI Model Is Best. The Embedding Decides Whether It Works

Choosing the language model took a day. Getting the embeddings right took three weeks. In any AI that answers from your documents, the embedding decides what the model is even allowed to see.
Production AI on a Single 32GB GPU: The VRAM Budget Nobody Shows You

One machine, multiple clients, no cloud. Exactly how the VRAM breaks down on a 32GB GPU running production B2B AI, plus two lessons that cost us hours.
Stop managing manuals, start managing workflows

Most leaders agree that AI will not fix a broken process. True. But this is exactly why SMEs should use AI as the forcing function to choose one way of working, encode it, and remove the daily dependence on tacit knowledge. Policies live in PDFs, work lives in people’s heads. New joiners copy Bob or […]