Chunking strategy comparison
This is the comparison table from the full chunking guide, kept here for quick reference. It maps the chunker names you will see across vendors, shows how boundaries are chosen, and summarizes the most common tradeoffs.
| Complexity | Chunker strategy | Names you'll see | How boundaries are chosen | Retrieval unit | Typical knobs | Advantages | Disadvantages / common failure modes | Good fit |
|---|---|---|---|---|---|---|---|---|
| 0 | No chunking | "No chunking", "whole document", "document-as-unit" | Don't split; store each record or document as-is | Whole doc or record | Mostly retrieval knobs such as top-k and filters | Simplest; preserves full context; minimal ingest logic | Coarse embeddings; lost-in-the-middle risk; higher latency or cost if you stuff whole docs into prompts | FAQs, short tickets, short posts, already atomic content |
| 1 | Fixed-size (tokens or chars) | Weaviate fixed-size chunking, Pinecone fixed-size chunking, Chonkie TokenChunker | Cut every N tokens or characters regardless of meaning | Chunk | chunk_size, overlap, chunk_overlap | Fast, cheap, deterministic baseline; easy to A/B test | Breaks sentences and ideas; can orphan facts at boundaries | Quick baseline; messy text; speed-first ingestion |
| 2 | Sliding window | Sliding window, overlap chunking, windowed chunking, Chonkie Slumber Chunker | Like fixed-size, but step size is smaller than chunk size so chunks heavily overlap | Chunk | overlap ratio, step size | Reduces boundary loss without smarter semantics | Redundant index; higher embedding and storage cost; duplicate-heavy retrieval | When boundary misses hurt but you still need a deterministic path |
| 3 | Sentence-based | Sentence splitting, paragraph splitting, Chonkie SentenceChunker | Split on sentence boundaries and then pack to a size limit | Chunk | tokenizer, max tokens, overlap, min sentences | Cleaner thought units; fewer mid-sentence breaks | Sentence length varies; a single sentence can still be too context-poor | QA and summarization where sentence integrity matters |
| 4 | Recursive delimiter chunking | Recursive chunking, RecursiveCharacterTextSplitter, Chonkie RecursiveChunker | Try paragraph breaks first, then sentence breaks, then words until the chunk fits | Chunk | separator list, max size, min chunk size | Solid default; respects common structure more than fixed-size | Still heuristic; weird formatting can defeat it; still returns broken text | Articles, blog posts, papers, reports |
| 5 | Document-structure aware | Document-based chunking, header-based chunking, element packing, Unstructured smart chunking | Parse the format and split on structure such as headers, tags, pages, or semantic elements | Chunk | max_characters, new_after_n_chars, section rules | Better topical coherence; less random tearing; robust across file types when partitioning works | Depends on extraction quality; bad title detection can create tiny or wrong chunks | PDFs, HTML, Markdown, manuals, policies |
| 6 | Table-aware | Table chunking, specialized table chunkers | Identify tables and keep them isolated; split oversized tables separately | Table chunk | size limit, table serialization | Preserves row and column structure | Large tables can dominate retrieval and prompt budget | Financials, specs, legal tables |
| 7 | Code-aware | Code chunking, AST chunkers | Parse code syntax and split by logical code blocks such as functions, classes, or modules | Code chunk | language detection, chunk size, docstring or import preservation | Preserves runnable units; improves code retrieval | Cross-file dependencies remain hard; large functions still need splitting | Repos, SDK docs, notebooks, code-heavy knowledge bases |
| 8 | Semantic similarity chunking | Semantic chunking, context-aware chunking, Unstructured by_similarity, Chonkie SemanticChunker | Embed local sentences or elements and place boundaries where similarity drops or merge only if similarity remains high | Chunk | embedding model, similarity threshold, window size, max size | Better topical purity; fewer mixed-topic chunks | More ingest-time compute; brittle if topic drift is gradual | Dense legal, academic, or technical narrative text |
| 9 | Neural boundary detection | Neural chunking, learned boundary detection | A learned model predicts good boundaries from semantic-coherence patterns | Chunk | model choice, min chunk size | Less hand-tuning; can outperform naive separator lists | Opaque decisions; domain mismatch can create odd splits | Mixed corpora where heuristics underperform |
| 10 | LLM-based chunking | LLM-based chunking, proposition extraction | Ask an LLM to propose boundaries or rewrite text into retrieval-friendly units | Chunk | prompts, target size, cost and latency budget | High semantic quality on complex text | Slow, expensive, nondeterministic; risk of summary drift | High-value docs where quality beats cost |
| 11 | Agentic chunking | Agentic chunking, strategy-selection chunking | An agent inspects each document and chooses a strategy or mix, possibly adding metadata | Chunk | model choice, allowed tools, budget | Flexible across wildly different document types | Highest complexity; inconsistent across runs without strong monitoring | Regulatory and multi-format enterprise corpora |
| 12 | Adaptive chunking | Adaptive chunking | Keep one approach but dynamically adjust size or overlap depending on content density or structure | Chunk | dynamic size ranges, density rules | Avoids one-size-fits-all behavior within a document | Harder to predict index stats and reproduce exactly | Long docs with mixed dense and sparse sections |
| 13 | Late chunking | Late chunking, embed-first-split-later chunking | Embed the whole document with full context and derive chunk embeddings afterward | Chunk with document-aware embeddings | long-context embedding model, split rules, pooling method | Reduces context loss in embeddings; improves cross-reference understanding | Requires long-context embeddings and more compute | Technical or legal docs with cross-references |
| 14 | Hierarchical chunking | Hierarchical chunking, parent-child chunking, multi-level chunking | Create multiple layers of chunks and retrieve coarse-to-fine | Multi-level chunks | levels, sizes per level, parent-child retrieval logic | Answers broad and specific questions more gracefully | More indexing and retrieval logic; still breaks text | Textbooks, handbooks, manuals, long contracts |
| 15 | POMA chunksets plus cheatsheets | Chunksets, hierarchical chunksets, breadcrumb context, cheatsheets | Parse sentence-by-sentence, infer hierarchy, and group sentences into unbreakable root-to-leaf paths | Chunkset, later compiled into a cheatsheet | Minimal exposed knobs; relies on structure inference | Facts arrive with contextual lineage attached; structured, token-efficient context | Requires an integrated ingestion and retrieval design; different mental model from standard chunk retrieval | High-stakes grounding in policies, legal, compliance, and manuals |
For the narrative framing around this table, go back to the RAG chunking guide.
Continue reading
- Strategy landscape — detailed explanation of each strategy
- Common failure modes — why most strategies fail in similar ways
- POMA chunksets — the non-breaking alternative
- The full chunking guide — complete narrative guide
- PrimeCut product page — how POMA implements hierarchical chunking
- Pricing — PrimeCut Adaptive: 1 credit per page, minute, or 1,000 tokens extracted (whichever is more); 1 credit = €0.01, easy jobs cost less