🤖 AI Engineering · RAG Pipelines

Design a chunking strategy for retrieval

Choose chunk sizes, overlaps, and structure-aware splits so retrieved context actually helps the model answer.

foundation~25 minAI EngineersML EngineersLLM App Developers

Steps

1Profile source documents: length distribution, headings, tables, code blocks
2Pick structural splitter first (headings/paragraphs) over fixed character cuts
3Set chunk size from your model's effective context budget minus prompt overhead
4Add 10-20% overlap so facts spanning boundaries survive retrieval
5Attach metadata per chunk: source title, section path, date, permissions
6Build 20 golden questions and measure retrieval hit-rate before scaling up

Common Pitfalls

▲Fixed-size chunks that slice tables and lists into meaningless fragments
▲Chunks larger than half the context window, starving the actual answer space
▲No overlap, causing answers that cite half a sentence cut at a boundary

Success Signals

✓Recall@5 above 85% on golden set
✓Median chunks-per-answer under 6

Commands

Install with skills CLI

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-chunking-strategy

Install globally

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-chunking-strategy -g

Tags

#rag#embeddings#retrieval#ai-engineering#rag-pipelines

Install this skill

$ npx skills add aniruddhaadak80/skills --skill rag-pipelines-chunking-strategy -a opencode

SKILL.md

# Design a chunking strategy for retrieval

> Choose chunk sizes, overlaps, and structure-aware splits so retrieved context actually helps the model answer.

## When to Use This Skill

Choose chunk sizes, overlaps, and structure-aware splits so retrieved context actually helps the model answer.

## Steps

1. Profile source documents: length distribution, headings, tables, code blocks
2. Pick structural splitter first (headings/paragraphs) over fixed character cuts
3. Set chunk size from your model's effective context budget minus prompt overhead
4. Add 10-20% overlap so facts spanning boundaries survive retrieval
5. Attach metadata per chunk: source title, section path, date, permissions
6. Build 20 golden questions and measure retrieval hit-rate before scaling up

## Common Pitfalls

- Fixed-size chunks that slice tables and lists into meaningless fragments
- Chunks larger than half the context window, starving the actual answer space
- No overlap, causing answers that cite half a sentence cut at a boundary

View source on GitHub ↗

Combine BM25-style lexical search with dense embeddings and fuse results so both rare terms and paraphrases are found.

🤖 AI Engineering·~35m

Open →

Evaluate RAG answer quality automaticallyadvanced

Score groundedness, relevance, and completeness with judge prompts plus deterministic checks wired into CI.

🤖 AI Engineering·~45m