Practice question from Foundations of Modern AI Search
A customer support RAG system chunks a 2000-word troubleshooting guide into 512-token segments with no overlap. Users searching for "connection timeout errors" receive chunks that mention timeouts but lack the diagnostic steps that appear 50 tokens later in the next chunk. What chunking modification would most directly solve this problem? A. Reduce chunk size to 256 tokens to improve retrieval precision B. Increase chunk size to 1024 tokens to capture more context per chunk C. Implement 50-100 token overlap between consecutive chunks D. Switch to semantic chunking that respects section boundaries in the document E. Use sentence-window retrieval to expand context after initial matching
Answer
C
Explanation
Chunk overlap ensures that information near boundaries appears in multiple chunks, preventing the exact problem described where diagnostic steps following error mentions get separated. With 50-100 token overlap, the timeout mention and its solution would both appear in at least one chunk. Option A would worsen the problem by creating more boundaries. Option B might help but increases irrelevant context and reduces precision. Option D is good long-term practice but doesn't specifically address boundary splitting. Option E adds complexity and requires a second retrieval step, while overlap directly prevents the split in the first place.