Practice question from Foundations of Modern AI Search
A pharmaceutical company's drug interaction database uses 512-token chunks. A query about "contraindications for warfarin in elderly patients with renal impairment" retrieves chunks mentioning warfarin and elderly patients separately, but misses the complete interaction profile that spans 800 tokens. What indicates this is specifically a chunking problem rather than an embedding quality issue? A. The relevant information exists in the database but is split across multiple chunks that don't individually match the query well B. The embedding model doesn't understand medical terminology like "warfarin" or "contraindications" C. The vector database is using the wrong similarity metric for pharmaceutical queries D. The retrieval is returning too few results to capture all relevant information E. The prompt engineering doesn't instruct the LLM to synthesize information from multiple sources
Answer
A
Explanation
The diagnostic sign of a chunking problem is that the complete relevant information exists but is fragmented across chunks in a way that prevents effective retrieval. The 800-token interaction profile split into 512-token chunks loses semantic coherence, and individual fragments don't match the multi-faceted query as well as the complete profile would. Option B would cause all warfarin queries to fail, not just complex ones. Option C would affect all retrievals systematically. Option D might help but doesn't address the root cause of information fragmentation. Option E is a generation problem, not a retrieval problem.