AI Foundations: RAG & Vector Search Practice Questions

Pinecone (A) is a fully managed, cloud-native vector database with built-in metadata filtering and a large community following in the AI/ML space. Weaviate (B) offers a managed cloud tier (Weaviate Cloud Services), supports metadata filtering via its 'where' filter, and has a strong open-source community. SQLite (C) is a relational database with no native vector indexing or similarity search capability. Standard Redis (D) without the RediSearch or RedisAI modules does not support vector similarity search or metadata filtering. A flat file JSON store (E) has no query engine, indexing, or similarity search capability.

Options B and C are correct. B: BM25 is a sparse retrieval algorithm that scores based on TF-IDF-like statistics, giving higher weight to rare but relevant query terms. C: Because BM25 operates on exact term matches, it reliably retrieves documents containing specific product codes or rare terms that dense embeddings may not capture. Option A is incorrect; BM25 uses sparse inverted-index representations, not dense vectors. Option D is incorrect; BM25 runs efficiently on CPU using inverted indexes and does not require GPU acceleration. Option E is incorrect; BM25 has no semantic understanding — 'automobile' and 'car' would be treated as completely different tokens.

Option B is correct: 'stuffing' (or context stuffing) refers to inserting retrieved document chunks directly into the prompt context before generation. Option A (reranking) is a post-retrieval step that reorders candidates by relevance — it does not describe how context is placed in a prompt. Option D (HyDE) is a technique where a hypothetical answer is generated to improve retrieval query quality. Option C (hybrid retrieval) describes combining dense and sparse retrieval methods, not prompt construction.

Options A and D are correct. A: Parallelizing the multiple vector store queries means total retrieval time is bounded by the slowest single query rather than the sum of all queries — a direct and significant latency reduction. D: Caching results for common or structurally similar sub-queries (e.g., time-period or company-name variants) avoids redundant vector store round-trips, cutting latency for repeated patterns common in financial analytics. Option B is incorrect; increasing to 10 sub-queries multiplies cost and latency substantially and does not reduce the need for reranking — it increases it. Option C is incorrect; a bi-encoder reranker is effectively another retrieval/scoring pass and adds latency rather than reducing it; deduplication alone would be cheaper. Option E is incorrect; using GPT-4o for sub-query generation is already likely the setup, and it does not reduce the number of API calls — it keeps them the same while increasing per-call cost compared to a smaller model.

Context stuffing (also called prompt stuffing) is the technique of inserting all retrieved passages directly into the prompt context so the LLM can reference them when generating a response. (A) is wrong — prompt chaining refers to sequencing multiple LLM calls where the output of one feeds into the next. (B) is wrong — fine-tuning injection is not a real RAG term; fine-tuning involves updating model weights. (C) is wrong — retrieval fusion is a technique for combining results from multiple retrieval strategies (e.g., BM25 + semantic), not for constructing the final prompt.