Best Chunking Strategies for RAG in 2026: A Practical Selection Guide
TL;DR
There is no universally best RAG chunking strategy. The correct method depends on document structure, question type, embedding behavior, reranking, and the context the generator needs.
Structure-aware section chunking is the best default for well-formed documentation. It preserves headings and semantic units while avoiding the cost of a model-driven splitter.
Recursive token chunking is the strongest baseline. It is simple, deterministic, and useful for measuring whether a more complex method earns its added ingestion cost.
Parent-child retrieval works better when precise matching and broad answer context conflict. Retrieve a small child, then return its larger parent to the model.
Page-level chunking is worth testing for page-oriented PDFs. NVIDIA's 2025 benchmark found it strongest on average across its tested datasets, but results still varied by corpus and query.
Evaluate retrieval and answer quality together. A strategy that raises vector recall can still hurt answers by adding repeated or irrelevant context.
What makes a chunking strategy “best” for RAG
The best strategy returns the smallest evidence unit that answers the query while preserving enough surrounding context to interpret it. Chunking happens after acquisition and cleaning but before embedding and indexing. Nstproxy Crawl can supply normalized Markdown or other selected outputs from authorized sites; it does not choose how those documents should be chunked.
A chunk boundary changes four things at once: what gets embedded, what a query can retrieve, how much unrelated text reaches the reranker, and what context the generator sees. That is why copying a chunk size from a framework tutorial rarely survives contact with real documents.
This guide compares each strategy with the same decision fields:
Does the method use tokens, paragraphs, headings, pages, or semantic change?
Best corpus
Which document structures make the signal trustworthy?
Query fit
Does it favor exact facts, multi-sentence explanations, or cross-section reasoning?
Context recovery
Can the system expand from a precise hit to a larger parent?
Ingestion cost
Does the method require parsing, embeddings, or an LLM during indexing?
Main failure
What information is most likely to be split, duplicated, or buried?
The RAG glossary provides the broader retrieval-and-generation context. Chunking is one design variable inside that system, not a quality switch by itself.
Strategy comparison
Strategy
Boundary signal
Best corpus
Query fit
Context recovery
Ingestion cost
Main failure
Structure-aware sections
Headings, lists, paragraphs
Documentation, articles, policies
Explanatory questions
Optional parent section
Low to medium
Bad or missing markup
Recursive token chunks
Ordered separators plus token limit
Mixed text with weak structure
General baseline
Overlap only by default
Low
Arbitrary semantic cuts
Parent-child
Small child plus larger parent
Long structured documents
Exact retrieval with broad answers
Built in
Medium
Parent adds excess context
Page-level
PDF page boundary
Reports, filings, manuals
Page-local facts and analysis
Adjacent pages if added
Low
Meaning crosses pages
Semantic
Embedding similarity changes
Transcripts and topic-dense prose
Topic-oriented questions
Optional window expansion
High
Threshold instability and cost
1. Structure-aware section chunking: best default
Structure-aware chunking splits on headings, paragraphs, lists, tables, and other document elements, then combines small adjacent elements up to a budget. It is the best default for technical documentation, knowledge bases, and web content whose structure carries meaning.
The Databricks RAG data-pipeline guidance distinguishes fixed-size, paragraph, format-specific, and semantic approaches and stresses that the optimum depends on the data and use case. Structure-aware splitting uses information the publisher already supplied, so it often preserves a section's question-and-answer relationship without another model call.
Keep the heading path as metadata, such as Product > Authentication > Token rotation. Repeat that path in the embedded representation when short paragraphs are ambiguous, but retain the clean text separately for final context. Tables and code blocks should be treated as atomic elements or converted into a representation designed for retrieval.
The main failure appears when HTML is template-heavy, headings are missing, or a PDF parser loses reading order. Clean the document first. A thoughtful web index pipeline should preserve canonical URL, title, heading path, update time, and content hash alongside each chunk.
2. Recursive token chunking: best baseline
Recursive chunking tries preferred separators in order—sections, paragraphs, sentences, then smaller units—until each chunk fits a token budget. It is the best baseline because it is deterministic, inexpensive, and available in most RAG libraries.
Use tokens rather than characters when the embedding model enforces a token limit. Start with a moderate size and modest overlap, then tune using real questions. The Azure AI Search chunking guide presents 512 tokens with 25% overlap as a starting point, not a universal optimum. That distinction is important: a support FAQ, source-code repository, and annual report do not share an ideal window.
Overlap can preserve sentences across boundaries, but it also duplicates vectors, increases storage, and may cause the same passage to occupy several top results. Deduplicate or diversify after retrieval. If large overlap is required to make answers coherent, the splitter is probably ignoring useful structure.
3. Parent-child chunking: best for precise retrieval plus context
Parent-child chunking indexes small child passages but maps each one to a larger parent section. The retriever matches the focused child; the application sends the parent—or a bounded region around the child—to the generator. This is also called small-to-big retrieval.
The Microsoft advanced RAG guide describes hierarchical indexes and Small2Big context expansion. The pattern works well when users ask for a precise clause but the answer requires definitions, exceptions, or steps around it.
Parent-child retrieval needs stable document and offset identifiers. Store the parent ID, child position, heading path, and content version. After retrieval, merge duplicate parents and preserve the most relevant child location so the generator does not receive the same section repeatedly.
The trade-off is context inflation. A small child can match one keyword while its parent includes several unrelated subsections. Limit the parent boundary to a coherent section, use a reranker on children, and compare full-parent retrieval with a fixed surrounding window.
Start RAG Chunking With Cleaner Documents
Use Nstproxy Crawl to collect bounded, structured website content before parsing, chunking, embedding, and retrieval.
4. Page-level chunking: best for page-oriented PDFs
Page-level chunking keeps each PDF page as one retrieval unit. It works when page boundaries matter to readers, tables or charts are page-local, and citations must reference a page number. Reports, filings, slide-like documents, and manuals are strong candidates.
In a 2025 evaluation, the NVIDIA RAG chunking benchmark reported the highest average end-to-end accuracy for page-level chunking across its tested datasets. The same evaluation also found that the optimum varied by dataset and query: some fact-oriented sets favored smaller chunks, while analytical questions benefited from larger or page-aligned context. The responsible conclusion is to test page chunks, not to declare pages universally superior.
Pages fail when a sentence, table, or section crosses a page break. Header and footer repetition can also dominate similarity. Remove repeated furniture, retain the page number as metadata, and test adding adjacent pages only after a hit rather than embedding multi-page windows by default.
5. Semantic chunking: best for topic shifts without markup
Semantic chunking detects changes in meaning, often by comparing embeddings for neighboring sentences or groups. It can help with transcripts, meeting notes, or long prose where heading structure is absent and topic boundaries matter more than length.
Its appeal is also its risk. Results depend on sentence segmentation, embedding model, similarity threshold, and the window used to smooth local changes. Changing the embedding model can silently change the corpus boundaries and invalidate cached evaluations. Semantic splitting also adds embedding work during ingestion before the final chunks are embedded again.
Use it only after the recursive baseline has a documented failure, such as repeatedly mixing adjacent topics or separating a discussion from its conclusion. Version the splitter configuration and retain source offsets so chunks can be reproduced. For noisy conversational text, compare semantic boundaries with speaker turns; the cheaper signal may be equally useful.
How to choose by document and query type
Use document structure as the first routing signal, then test against query behavior:
Product documentation: structure-aware sections, with parent-child expansion for long topics.
Short support articles: paragraph or recursive chunks with little overlap.
Financial reports and manuals: page-level plus section metadata; test table-aware parsing separately.
Transcripts: speaker-turn or semantic chunks, with timecodes retained.
Source code: syntax-tree or symbol boundaries rather than prose strategies.
Mixed enterprise corpus: route by MIME type and parser output instead of forcing one global splitter.
For fresh website corpora, an automated data collection pipeline should recrawl only what policy permits, detect content changes, and re-embed affected chunks. Rebuilding every vector on every run wastes resources and complicates traceability.
An evaluation protocol that reveals real differences
Start with a question set sampled from actual usage. Include exact lookups, explanatory questions, multi-part requests, negative questions, and queries whose answer spans a boundary. Label the supporting source passages before tuning the splitter.
Measure at least four layers:
Retrieval recall: did any top result contain the required evidence?
Context precision: how much retrieved text was relevant to the question?
Answer faithfulness: were answer claims supported by supplied context?
Operational cost: how many chunks, embedding tokens, reranker inputs, and generation tokens were required?
Keep the embedding model, retrieval method, reranker, prompt, and generator fixed while comparing chunking. Then tune the winning strategy jointly with top-k and context assembly. Otherwise a stronger reranker may be mistaken for a stronger splitter.
Inspect failures, not only averages. One method may perform well overall while consistently losing table headers, exceptions, or cross-page definitions. Those errors affect product decisions more than a small aggregate improvement.
Prepare web content before chunking it
Chunking cannot repair a broken input document. Navigation, cookie banners, duplicate mobile markup, missing JavaScript-rendered content, and malformed reading order become low-quality vectors regardless of the splitter.
Nstproxy Crawl can collect authorized sites with page, depth, include, and exclude controls and return formats suited to downstream processing. Use Markdown when headings and prose matter; retain HTML when tables, attributes, or structural parsing are important. The AI search agent guide shows where this ingestion layer fits before retrieval and synthesis.
Final verdict: use structure first, then prove exceptions
Structure-aware section chunking is the best general default, while recursive token chunking is the benchmark every more complex method should beat. Use parent-child retrieval when precise hits need wider context, test page-level chunks for page-oriented PDFs, and reserve semantic splitting for corpora where topic changes are real but markup is weak.
The next step is to label a representative question set and run the same retrieval-and-answer evaluation across two or three strategies. If the source pages are incomplete or noisy, fix acquisition with Nstproxy Crawl before spending time on splitter thresholds.
Give your RAG pipeline cleaner source documents
Use Nstproxy Crawl to produce bounded, structured web inputs with retained URLs and selectable formats, then evaluate chunking strategies over a reproducible corpus.
There is no universal best chunk size. Start with a moderate token window, then tune it against representative queries while holding the embedding model, retriever, reranker, and generator constant.
Q: How much overlap should RAG chunks use?
Use only enough overlap to preserve meaning across unavoidable boundaries. High overlap increases storage and duplicate retrieval; structure-aware or parent-child approaches often preserve context more efficiently.
Q: Is semantic chunking better than fixed-size chunking?
Semantic chunking is better only when its topic boundaries improve measured retrieval and answer quality enough to justify extra ingestion work. Recursive fixed-token chunking remains the more reproducible baseline.
Q: Should PDFs be chunked by page?
Page chunking is a strong candidate for page-oriented reports, manuals, and filings, especially when page citations matter. It performs poorly when key sentences, tables, or sections routinely cross page boundaries.
Marcus Chen
Aug. 26th 2026
110M+ real IPs with 99.9% access success
Blazing-fast average response ~0.5s for high-concurrency tasks
From only $0.1/GB
Get immediate access to premium residential, datacenter, IPv6 and ISP proxy pools.
