RAG Chunking Strategy Designer

When to use this

Use when setting up or re-tuning a document ingestion pipeline for retrieval-augmented generation.

Prompt template

You are a RAG ingestion specialist. Document type: {{document_type}}.
Typical structure notes: {{structure_notes}}
Retrieval goal: {{retrieval_goal}}
Embedding model context window (tokens): {{context_window}}

Recommend:
1. A chunk size (tokens) and overlap, with the reasoning.
2. A splitting rule (by heading, paragraph, sentence, fixed-size, or a hybrid), and why it fits this document type.
3. What metadata to attach to each chunk (e.g. source, heading path, page number) and why it helps retrieval or citation.
4. One risk this strategy does NOT solve, and what would.

Customize

Preview

You are a RAG ingestion specialist. Document type: {{document_type}} missing.
Typical structure notes: 
Retrieval goal: {{retrieval_goal}} missing
Embedding model context window (tokens): 8192

Recommend:
1. A chunk size (tokens) and overlap, with the reasoning.
2. A splitting rule (by heading, paragraph, sentence, fixed-size, or a hybrid), and why it fits this document type.
3. What metadata to attach to each chunk (e.g. source, heading path, page number) and why it helps retrieval or citation.
4. One risk this strategy does NOT solve, and what would.

Usage instructions

Fill in your document type and what you need retrieval to be good at (exact clause lookup vs. broad thematic summarization); the answer differs a lot between the two.

Example output

A recommendation for heading-based chunking (~300-500 tokens, no overlap needed since headings are natural boundaries), metadata to attach (endpoint name, HTTP method), and a flagged risk (cross-references to other endpoints won't be resolved by chunking alone).

Limitations

This is a starting heuristic; validate with the evaluation plan from the companion RAG Evaluation Plan prompt rather than assuming it's correct.

Rate this prompt