Parent-Child Retrieval

Introduction

Parent-Child Retrieval is a retrieval strategy used when a small chunk is good for finding the correct information, but does not contain enough surrounding context for the LLM to answer correctly.

The main idea is:

Child Chunk  → used for retrieval
Parent Chunk → returned to the LLM

This is useful when production evaluation shows that the correct chunk is already appearing in the Top-K results, but the final answer is still incomplete because an exception, condition, or related rule exists nearby.

---

How It Works

Imagine we have API documentation:

POST /payments

Transactions above $10,000 require manual verification.

Enterprise accounts using batch settlement cannot use this endpoint.
They must use /batch-payments instead.

We can first create a larger parent chunk, then split it into smaller child chunks:

Parent: POST /payments
│
├── Child 1: endpoint description
├── Child 2: $10,000 verification rule
└── Child 3: enterprise batch settlement restriction

Suppose a user asks:

Can an enterprise batch-settlement account send a $15,000 payment?

Semantic search may correctly retrieve:

Child 2:
Transactions above $10,000 require manual verification.

If only this child is passed to the LLM, it may incorrectly answer:

Yes, but manual verification is required.

The problem is not that retrieval completely failed. It found the relevant location, but the retrieved context was too small.

With Parent-Child Retrieval:

Query
  ↓
Search small Child chunks
  ↓
Child 2 matched
  ↓
Find parent_id
  ↓
Retrieve POST /payments Parent
  ↓
Send larger Parent to LLM

Now the LLM also sees the enterprise restriction and can produce a more complete answer.

In production, I would test this when I repeatedly observe:

Correct information appears in Top-K
        ↓
Final answer is still incomplete
        ↓
Required information exists nearby

I could then A/B test:

Version A
Search 300-token children
→ Return children

Version B
Search the same 300-token children
→ Return their 1,200-token parents

Then compare answer correctness and completeness while monitoring token usage and latency.