Production AI Architecture Interview Suite

What is the fundamental difference between dense retrieval and sparse retrieval architectures?

How do 'chunk size' and 'chunk overlap' choices directly influence LLM generation behavior and resource constraints?

Why are dedicated spatial data indices required for vector search instead of traditional relational B-Trees?

What is the operational mechanism of a Cross-Encoder compared to a Bi-Encoder framework?

Define the concept of 'Embedding Collisions' and explain why they happen in practice.

What role does Cosine Similarity play relative to Dot Product or Euclidean Distance when evaluating normalized vector embeddings?

How does the 'Lost in the Middle' phenomenon impact multi-document RAG pipelines, and how do you mitigate it architecturally?

Compare the trade-offs between Pre-Filtering, Post-Filtering, and Native In-Query Filtering during metadata vector filtering.

Walk through the structural execution flow of Parent-Child Chunking strategies (Auto-Merging Retrieval).

How would you design an implementation strategy for Query Rewriting/Expansion to resolve ambiguous or conversational user inputs?

What is semantic drift within an active enterprise RAG system, and how do you monitor it?

Describe the operational advantages of implementing a hierarchical multi-stage retrieval pipeline.

How does contextual compression reduce noise in long documents pulled by vector search?

Your RAG system indexes docs updated 20+ times a day. Users get stale answers within hours. How do you keep the vector index fresh without re-embedding the entire corpus every time?

Compare the scale performance trade-offs of Hierarchical Navigable Small World (HNSW) graphs versus Inverted File Indexes (IVF) at 100M+ vector scale.

Explain GraphRAG architecture. How does building a knowledge graph over a vector index change multi-document global entity reasoning?

How do you design multi-tenant security structures at the vector database layer to prevent cross-organization document visibility?

Detail a method for handling multi-modal ingestion (tables, complex chart images) inside a production RAG system.

How do you address vector space mismatch when using different models for localized queries vs long-form historical enterprise text?

Explain the role of self-RAG loops where an LLM critiques its own retrieved context blocks before completing final token generation.

What is the primary role of a 'Parent Document Retriever' in a RAG pipeline, and how does it separate the search index from LLM synthesis text?

What are the structural trade-offs of using Hierarchical Chunking (e.g., recursive text splitting) vs Fixed-Size Chunking?

How does the selection of distance metrics (L2 Distance vs Inner Product vs Cosine) depend on the chosen embedding model?

Explain the implementation strategy for Metadata Enrichment (e.g., adding summary text or keywords to chunks) to boost retrieval accuracy.

How does the choice of tokenizers (e.g., TikToken vs HuggingFace Fast Tokenizers) affect runtime vector boundary parsing?

What is the operational purpose of Forward-Looking Active Retrieval (FLARE) in continuous long-form RAG workflows?

How do you systematically handle duplicate or highly redundant content within enterprise document index pipelines?

Describe the architectural execution steps of utilizing a re-ranking model over metadata filter returns.

How do you design an active partitioning engine inside a single vector index to accommodate custom enterprise document security clearance tiers?

Detail a cold-start vector index migration pattern to swap base embedding models (e.g., OpenAI text-embedding-3 to a local custom model) on a 50M+ document cluster without causing application downtime.

Explain how vector quantization techniques (Scalar Quantization [SQ] vs Product Quantization [PQ]) balance accuracy drops against RAM footprint reductions.

How do you design a real-time table extraction RAG framework for documents where data spans across nested canvas layouts or merged border lines?

What architectural parameters dictate whether to use a managed vector store instance vs a native vector plugin built directly into a relational database cluster (e.g., pgvector)?