RAG Engineering Mastery2 / 10
Chunking — The Decision That Sets Your Ceiling
You can't retrieve what you chunked badly. Chunking is the most under-rated lever in RAG — and the cheapest to get right.
Retrieval can only return the chunks you created. If a chunk splits an idea in half, no embedding model on earth will retrieve it whole. Chunking sets the ceiling on everything downstream.
Three strategies
- Fixed-size — split every N tokens with overlap. Simple, fast, dumb. Fine for uniform prose, bad for structured docs.
- Structural — split on the document's own boundaries: headings, sections, list items, code blocks. Respects meaning for free.
- Semantic — split where the topic shifts (embedding-distance based). Best quality, higher cost.
Start structural; it captures most of the win at near-zero cost.
Size and overlap
Too small and a chunk loses context; too big and retrieval gets noisy and the prompt gets expensive. A pragmatic default: 300–600 tokens with ~15% overlap, then tune against your eval set (article 7).
Metadata is the quiet superpower
Attach metadata to every chunk: source, title, section, date, URL. It powers filtered retrieval (only this product, only docs after this date) and lets the generator cite precisely.
Next: turning these chunks into vectors, and where to store them.
Series — RAG Engineering Mastery
- Part 01Why Naive RAG Fails in ProductionThe 50-line vector-search demo that wows in a notebook falls apart the moment real users ask real questions. Here is why — and the map out.
- Part 02Chunking — The Decision That Sets Your Ceiling — you are hereYou can't retrieve what you chunked badly. Chunking is the most under-rated lever in RAG — and the cheapest to get right.
- Part 03Embeddings & Vector Stores 101An embedding turns meaning into geometry. A vector store makes that geometry searchable in milliseconds. Get both right and retrieval gets easy.
- Part 04Hybrid Retrieval — Keyword + VectorVector search understands meaning but fumbles exact terms, IDs, and rare words. Keyword search nails those and misses paraphrase. Use both.
- Part 05Re-Ranking — The Cheap Quality WinRetrieval gets you 30 plausible chunks. A re-ranker reads them against the actual question and floats the truly relevant few to the top.
- Part 06Prompting the Generator — Grounding & CitationsGreat retrieval is wasted if the model ignores it or can't point to its sources. Grounding is a prompt-design discipline, not an afterthought.
- Part 07Evaluation — You Can't Improve What You Don't MeasureWithout an eval set, every RAG change is a vibe. With one, you tune chunking, retrieval and prompts with a number that tells you if you helped or hurt.
- Part 08Handling Hallucinations & GuardrailsWhen retrieval comes up empty, a helpful model invents. Guardrails turn 'confidently wrong' into 'honestly unsure' — the difference users actually trust.
- Part 09Cost & Latency DisciplineA RAG query touches embeddings, a vector DB, a re-ranker and an LLM. Each adds milliseconds and cents. At scale, discipline here is the difference between a margin and a bonfire.
- Part 10The Production RAG Reference ArchitectureEvery piece, assembled: ingestion, hybrid retrieval, re-ranking, grounded generation, guardrails, eval and caching — the blueprint you can ship.