Skip to main content

Chunking and search options

Chunking divides a document into searchable units called chunks. This page summarizes the criteria for selecting a chunking strategy, chunk size, embedding model, and search method.

Process a document with the displayed defaults first, then review the result under Review search results. Change options only when a required document is not retrieved or the returned chunk scope is unsuitable.

Chunking strategies

StrategyBehaviorDocuments to consider
HybridCombines document structure with fixed-size splitting.General documents that mix several formats
HierarchicalPreserves parent and child section structure.Long policies, papers, and specifications
Fixed sizeSplits at a configured length regardless of structure.Logs, transcripts, and unstructured text
Parent-childCreates small chunks for retrieval together with larger chunks for context.Documents that require both precise retrieval and broad context
SentenceSplits at sentence boundaries.Meeting minutes and question-and-answer records
SemanticSplits where meaning changes significantly between sentences.Long documents with frequent topic changes
AgenticAnalyzes paragraphs and groups semantically similar content.Long documents with inconsistent structure

Do not judge results from the strategy name alone. Process part of the same document with different strategies and compare the scope and order of search results.

Apply different chunking settings to an existing document

Existing documents are not processed again automatically

The chunking strategy and values are stored separately when each document is added. Applying new settings to another document does not change documents that have already been processed.

Reindex uses the settings already stored on that document. To split an existing document with different settings, delete it and add it again with the required settings. This differs from the full reset used to change the search method or embedding model.

Common chunking values

The on-screen defaults and allowed ranges are as follows.

OptionDefaultRangeEffect
Maximum chunk length500100–4,000A smaller value creates more precise retrieval units; a larger value includes more context in each result.
Overlap length500–500Sets the range repeated in adjacent chunks. A larger value also increases duplicated content.

The Agentic strategy uses its dedicated on-screen values instead of the common values.

Semantic strategy values

OptionDefaultRangeDescription
Embedding modelKoE5Displayed model listCalculates similarity between sentences.
Breakpoint typeGradientPercentile, standard deviation, interquartile range, gradientSets the calculation used to determine semantic boundaries.
Breakpoint threshold95.00–100Sets the boundary criterion for the selected breakpoint type.
Minimum chunk tokens5010–500Prevents excessively short chunks.

Breakpoint types produce different results by document. Compare chunk boundaries and search results created by each type before selecting one.

Agentic strategy values

OptionDefaultRangeDescription
Embedding modelKoE5Displayed model listCalculates similarity between paragraphs.
Maximum paragraph tokens25050–1,000Sets the maximum length of an initial paragraph to analyze.
Similarity threshold0.750–1Sets the criterion for grouping adjacent paragraphs.
Minimum chunk tokens6010–500Sets the minimum final chunk length.
Maximum chunk tokens51264–2,048Sets the maximum final chunk length.

Embedding model

An embedding model converts text to numeric vectors for semantic search. Available models depend on system configuration, so review the list on the knowledge creation screen or Settings tab.

Test the following with actual documents before choosing a model.

  • Confirm that the model supports the document's primary language.
  • Compare the order and scores of related documents for the same query.
  • Compare document-processing time and search response time.

More dimensions do not always produce more accurate results. Do not choose based only on model name or dimension count.

Change the embedding model

Changing the embedding model after adding documents requires clearing the existing documents and chunks. Review the reset procedure under Change knowledge settings first.

Search methods

MethodBehaviorSuitable queries
Semantic searchFinds document chunks with similar meaning.Natural-language questions and queries containing synonyms or similar expressions
Keyword searchFinds document chunks containing entered words.Product names, code, policy numbers, and exact terms
Hybrid searchCombines semantic and keyword search results.Questions that require both meaning and exact terms

The appropriate method depends on the documents and questions. Choose representative queries, then compare each method under Review search results.