Retrieval Is the Hard Part of Enterprise AI

Tomas NovakSep 8, 20261 min read2,210 reads

Most failed internal AI assistants fail at retrieval, not generation. The model was fine; it was handed the wrong context.

Fix the corpus first

Deduplicate, date-stamp and label ownership before indexing anything. Stale documents are worse than missing ones.

Score the retriever independently of the model. Teams that do this find the real bug in days.

Tomas Novak builds enterprise search and retrieval systems for regulated industries.

One weekly email with the best technology, business and finance stories from our contributors.

Frontier models are extraordinary generalists, but most production AI tasks are cheaper, faster and more reliable on compact models.

Hiroshi TanakaSep 19, 20263,416 reads

Comments are coming soon. In the meantime, share your take with us on social media.