AIMentor — AI Expertise Learning Platform

Untimed. Answers are graded when you submit, with explanations and spaced-repetition scheduling.

Question 1easy
Why is chunking often described as the highest-leverage knob in a RAG system?
Because smaller chunks are always cheaper to embedBecause retrieval can only ever return a chunk, so the answer must live cleanly inside oneBecause the LLM cannot read more than one chunk at a timeBecause chunking replaces the need for an embedding model
Question 2easy
You size chunks at '200 words'. Why can this silently break your token budget?
Words and tokens are the same thingEmbedding APIs and context windows are measured in tokens, and word→token ratio varies with the text200 words is always exactly 200 tokensWords are larger than tokens, so you always undershoot
Question 3easy
What does a tokenizer like tiktoken give you that a plain .split() does not?
The exact token ids the model uses, so chunk sizes match the real budgetFaster string splittingAutomatic removal of stop wordsA guarantee that chunks end on sentence boundaries
Question 4medium
In recursive character splitting, what is the ordering strategy for separators?
Random separator each pass to avoid biasSmallest boundary first (words), escalating to paragraphs only if neededLargest natural boundary first (paragraphs), falling back to smaller ones only when a piece is still over budgetAlways split on a fixed character count
Question 5medium
What is the main advantage of splitting on natural boundaries over a fixed token window?
It produces fewer chunks in every caseIt avoids cutting through the middle of sentences, tables, or Q&A pairs, so chunks stay coherentIt removes the need for overlap entirelyIt makes every chunk exactly the same size
Question 6medium
Which capabilities does attaching metadata (source, section, position) to chunks unlock? (Select all that apply)
Filtered retrieval (restrict the search to one document, section, or date range)Citations in the answer (e.g. 'source: pricing.md § Plans')Debugging which document a bad chunk came fromAutomatically increasing embedding accuracy
Select all that apply.
Question 7medium
When you filter retrieval by metadata (e.g. source='pricing.md'), when does the filter apply relative to similarity scoring?
After scoring — it re-ranks the top-k by metadataBefore scoring — it narrows the candidate pool, then similarity ranks what remainsIt replaces similarity scoring entirelyOnly at generation time, inside the prompt
Question 8hard
Parent-child (hierarchical) indexing resolves which tension?
Embedding cost vs. storage costSmall chunks match precisely but lack context, while large chunks carry context but retrieve impreciselyLatency vs. throughputOpen-source vs. proprietary embedding models
Question 9hard
In the 'small-to-big' parent-child pattern, what is embedded and what is handed to the generator?
Embed the parents; hand the children to the generatorEmbed the children; hand the matched children's parents to the generatorEmbed and generate from the same medium-sized chunkEmbed the metadata; hand the raw document to the generator
Question 10hard
When a single paragraph exceeds max_tokens, what does a well-built recursive splitter do? (Select all that apply)
Fall back to smaller separators (sentences, then words)Hard-cut on tokens only as a last resort when no separator remainsSilently drop the paragraphMerge it with the next document
Select all that apply.
Question 11medium
For an FAQ where each entry is a self-contained Q&A pair, what chunking approach usually works best?
Split every entry into 50-token pieces for precisionOne chunk per entry — the unit is already self-containedConcatenate all entries into one large chunkChunk purely by character count, ignoring entry boundaries
Question 12medium
How should you decide on a chunk size for a new corpus?
Pick 512 tokens because it is the standardUse whatever the first tutorial you read usedTry 2–3 configs and measure context recall against a golden set of question→answer pairsAlways use the smallest size the embedding model allows
Question 13medium
What unit should you size chunks in to reliably respect embedding and context-window limits? (one word)
Question 14hard
Name the indexing pattern that embeds small chunks for precise matching but returns larger surrounding chunks for context.

0 / 14 answered

Chunking & Indexing Strategies — Practice quiz