quiz · 4 min Retrieval & RAG Check What is the core reason RAG reduces hallucinations compared to pure parametric recall?The model conditions its answer on retrieved text rather than recalling facts from weightsThe model is retrained on new data at query timeRAG fine-tunes the model on retrieved chunks before answering, updating its weightsRetrieval increases the model's context window size permanentlyIn the tiny retriever built in Lesson 15, how does the score() function select the best document?It computes cosine similarity between embedding vectorsIt ranks documents by recency of their last updateIt counts how many query tokens appear in the document using set intersectionIt calls an external API to score relevanceAccording to the RAG failure modes lesson, what is the metric for measuring whether relevant chunks appear in the top-k results?BLEU scorePerplexityEmbedding vector dimensionalityrecall@kWhat does the RAG failure modes lesson say a well-designed RAG system should do when the retriever finds no relevant source?Fall back to parametric knowledge so the user always gets an answerReturn an explicit 'no supporting source' answer rather than guessingRe-run the query with a larger k until a source is foundExpand the chunk size until the index returns somethingThe 'lost in the middle' phenomenon described in Lesson 16 means that pasting in many retrieved chunks:Forces the model to cite every chunk it was givenPrevents stale context from reaching the modelImproves answer quality by giving the model more evidence to draw onOften produces worse answers because models reliably miss information in the middle of long contextsSubmit Mark complete ← RAG Failure Modes Why Agents Need Memory →