RAG & Vector Search: free practice questions

Correct: A cross-encoder reranker (such as Cohere Rerank) takes each (query, document) pair and scores them jointly, considering full interaction between query and document tokens. This is far more accurate than the initial bi-encoder similarity score and improves precision — ensuring the top-k chunks passed to the LLM are truly the most relevant. | Wrong A: Rerankers score existing embeddings or raw text; they do not regenerate embeddings. Re-embedding is a separate process. | Wrong B: Rerankers do not compress or summarize documents; they reorder the existing retrieved set. Summarization is a different technique (map-reduce chains, etc.). | Wrong D: Date filtering is a metadata pre-filter applied before or during retrieval, not a function of a cross-encoder reranker.

Context recall in RAGAS measures whether the retrieved context contains all the information needed to answer the question (i.e., coverage). A low score means relevant chunks exist in the corpus but are not being retrieved — the system has incomplete recall. The answers may still feel acceptable if the retrieved chunks happen to cover the most common question patterns, but coverage of edge cases is poor. (A) describes low faithfulness, not low context recall. (C) describes a precision problem — the reranker promoting bad results — which would lower faithfulness or answer relevance, not context recall. (D) is a red herring; dimensionality affects latency, not recall scores directly.

Correct: Bi-encoders (used in the initial retrieval) embed query and document independently and compare them via dot product — efficient but unable to model token-level interactions. Cross-encoders take the concatenation of query + document as input, enabling rich attention-based interactions between every query token and every document token. This yields much better relevance scoring at the cost of higher latency, making it ideal as a reranking step over a small candidate set. | Wrong B: Cross-encoders are significantly slower than bi-encoders because they require a forward pass for every query-document pair; they are never used for full-index search. | Wrong C: Cohere Rerank uses a neural cross-encoder model, not BM25; it is a learned relevance model, not a keyword scorer. | Wrong D: Model size correlates loosely with capability but is not the defining reason cross-encoders outperform bi-encoders here; the architectural difference (joint vs. independent encoding) is the key factor.

Correct: Qdrant is an open-source, self-hostable vector database written in Rust. It natively supports payload (metadata) filtering, hybrid search via sparse+dense vector combinations, and can be deployed on-premises or in private cloud environments — satisfying all three stated requirements. | Wrong A: Pinecone is a fully managed cloud service (SaaS) and cannot be deployed on-premises, which violates requirement 3. | Wrong C: MySQL with LIKE queries performs lexical pattern matching, not vector similarity search; it has no concept of dense vector retrieval or hybrid search and would be completely unsuitable. | Wrong D: OpenAI's embedding API is a model inference API that returns embedding vectors; it is not a vector database, has no persistent storage, no filtering, and no retrieval endpoint.

LLMs are next-token predictors trained on large text corpora; they generate plausible-sounding text based on learned patterns but have no internal fact-checking mechanism, which is the root cause of hallucination. Low temperature (A) actually makes outputs more deterministic and conservative, not more hallucinatory—high temperature is more likely to cause wild outputs, but neither is the root cause. The embedding model (C) is used for retrieval in RAG pipelines, not during standard generative inference, so it is unrelated. Tokenizer behavior (D) can affect output quality in edge cases but does not cause systematic factual hallucination.