Explain parent-child (small-to-big) retrieval and sentence-window retrieval.

30-second answerSay your answer out loud first, then reveal.

Small child chunks are searched with the query, the matching child's parent_id leads back to its parent section, and the whole parent section is sent to the LLM.

Parent-child (small-to-big)

Index small child chunks with a parent_id.
Search children; map hits to parents; deduplicate parents (several children may share one).
Send parents (or parent sections) to the LLM.

Sentence-window

Index individual sentences.
On a hit, return ±N sentences around it.

Benefits: precise matching and complete context; fixes "the answer chunk lacks context" problems.

Trade-offs

Bigger context per hit means more tokens. Limit the number of parents.
Extra storage and a lookup step.
If parents are too large, you're back to noisy context.

Related idea: multi-vector indexing. Store several representations per chunk (summary, hypothetical questions, raw text) that all point to the same source text.

PreviousHow do you fuse results in hybrid search? Explain Reciprocal Rank Fusion (RRF).

Little by little, you're building something great.