Advanced Retrieval-Augmented Generation (RAG) Quiz

1

In Retrieval-Augmented Generation, which retrieval approach has been shown to yield higher recall for long-tail entities in open-domain question answering compared to traditional BM25?

20 points

A Sparse TF-IDF weighting with BM25+ B Dense vector retrieval using cross-encoder fine-tuning C Hybrid retrieval combining sparse and dense embeddings via weighted voting D Phrase-level n-gram retrieval with Levenshtein distance

2

When integrating retrieval into a large language model, which fine-tuning strategy most effectively preserves the alignment between retrieved passages and generated responses?

20 points

A Retrieval-augmented fine-tuning with a separate encoder for context B Dual-encoder fine-tuning that jointly trains passage and query encoders C Prompt engineering that limits token length of retrieved passages D Post-hoc consistency filtering of generated text

3

Which evaluation metric is most suitable for quantifying the quality of retrieved passages in a RAG system, as opposed to traditional generation metrics like BLEU?

20 points

A Recall@k B ROUGE-L C METEOR D Perplexity

4

Considering latency and accuracy trade-offs in large-scale RAG deployments, which technique provides the best balance for real-time applications?

20 points

A Fine-grained Faiss IVF PQ index over compressed embeddings B Exact nearest neighbor search using HNSW graph C Pre-filtering candidates with BM25 followed by dense re-ranking D Using a lightweight transformer encoder for retrieval

5

To mitigate hallucinations caused by retrieved content in RAG, which approach has empirical evidence for reducing hallucinated facts while preserving fluency?

20 points

A Retrieval grounding with consistency checks during decoding B Strictly limiting the number of retrieved passages to one C Adding a hallucination penalty term to the loss function D Relying solely on post-hoc filtering of generated text

Advanced Retrieval-Augmented Generation (RAG) Quiz

Quiz Info

Questions

Sitemap

Support

Contact Us