A retrieval system you cannot score is a retrieval system you cannot maintain
Why the evaluation harness has to be built alongside the retriever, not after it, and what to put in it.
Notes
Written when wehave somethingspecific to say.
Two posts is better than ten padded out to look like a content strategy.
What makes Bangla hard for systems built on English assumptions
Script, tokenisation, code-switching and OCR, and where each one breaks a pipeline that was not designed for it.