Knowledge Retrieval

How a model answers from material it was never trained on: cutting documents up so they can be found, searching by meaning, and telling whether the search or the answer is what went wrong.

1. 01Why a model needs to look things upThree reasons a model cannot answer from memory alone: a training cutoff, knowledge it never saw, and a habit of guessing confidently.about 6 minutes
2. 02Look it up, then answerThe basic loop: a question, a search, passages put in front of the model, an answer. Neither pasting a document in yourself nor retraining the model.about 7 minutes
3. 03Cutting documents into piecesWhy a store holds chunks rather than whole files, what goes wrong when they are too big or too small, and keeping context attached.about 7 minutes
4. 04Searching by meaningSearching a map where things that mean similar things sit close together, how that differs from matching words, and why systems often run both.about 8 minutes
5. 05Choosing what the model seesHow many passages to include and what ranking is doing, and why a longer stack can produce a worse answer than a short one.about 7 minutes
6. 06Answering from the sourcesTelling the model to answer from what it was given, asking it to point at the passage, letting it say the documents do not cover the question, and handling sources that disagree.about 7 minutes
7. 07When retrieval goes wrongThe wrong chunk, a missing or stale document, and the right passage ignored. The question that sorts them: did the search fail, or did the answer fail?about 8 minutes
8. 08Retrieval inside an agentWhat changes when the agent decides whether to search, rewrites the question and searches again, and when building a store buys nothing.about 8 minutes

Opening a lesson needs membership. During the preview it's free for invited members. See plans