A parent document retriever for chunks too small to hold their own context

Your chunks are small enough to embed well. The model still misses half a fact because the number landed in the next chunk.

AdvancedRead 8 minDoing: an afternoon, with edit access to your indexing code and wherever the retriever's lookup lives8 steps

▶ Start

What you'll have at the end

A retriever returning the full parent paragraph for a fact your old chunking used to split in half

Before you start

You need

A RAG pipeline that already retrieves and answers from your own documents through one flat small-chunk index, plus one real question it answers wrong or half right because the fact it needed was cut across two of those chunks.

Not covered

Choosing your original chunk splitter's boundary rule, and building the reranker or citation trail that runs on whatever the search ends up returning.

Leans on

8 steps

Where to go next

Split your documents into chunks that help RAG give better answers

start there first if your base splitter still cuts a document in an obviously wrong place; this recipe adds a second, bigger layer on top of whatever split you land on

Your RAG's retrieval hit rate has never been measured

measure whether swapping in full parents actually raised how often the right fact comes back, instead of trusting the one question this recipe fixed

Checked 18 Aug 2026

Part of the RAG (Retrieval Augmented Generation) cookbook