Parent Document Retrieval — AI Glossary

Reference details

Related terms

• Sentence Window Retrieval
• Chunk Overlap
• Text Splitter

How it works

1. Documents are split twice: into small child chunks for matching, and into larger parent units for reading.
2. Only the child chunks are embedded and indexed, so similarity is computed over short, focused passages.
3. Each child stores a reference to its parent, which is the link the whole technique rests on.
4. At query time children are retrieved, then deduplicated by parent so one parent is not fetched repeatedly.
5. The parents are what enter the prompt, so the model reads a coherent section rather than the fragment that matched.

Example: Precision at match time, context at read time

A 2,000-token section split into 200-token children gives 10 children. A query matching one child retrieves it and fetches the whole 2,000-token parent — a tenfold expansion from match to context. Embedding 10 short vectors instead of 1 long one also sharpens matching, because a 200-token chunk about one thing has a cleaner vector than a 2,000-token section about five. The cost is prompt size: three distinct parents is 6,000 tokens where three children would have been 600.

Commonly confused with

Both this and sentence-window retrieval index something small and return something larger, to get precise matching with usable context. They differ in what the larger unit is: here it is a pre-existing parent — a section or whole document — so it needs a document hierarchy to exist, and returns a unit an author defined. Sentence-window computes its span around the hit instead.

When to use it

Reach for it when:

• Matching is accurate but the retrieved fragment is too small for the model to answer from.
• Documents have real parent structure — sections, chapters, tickets — worth returning whole.
• Answers need surrounding context that the matching sentence alone does not carry.

Reach for something else when:

• The context budget cannot absorb several large parents.
• Documents have no meaningful parent unit, making the larger span arbitrary.
• Child chunks already contain complete answers, where the expansion is pure cost.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Parent Document Retrieval — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path