Skip to main content

Chunking Strategies

Chunking strategies determine how documents are split into smaller pieces for retrieval. Choosing the right chunking strategy can significantly improve retrieval quality.

πŸ“Š Strategy Overview​

StrategyBest ForDescription
Smart ChunkingGeneral documentsAuto-detect document structure
Sentence-basedPrecise retrievalSplit by sentence boundaries
SemanticComplex documentsSplit by semantic similarity

🧠 Smart Chunking​

Smart chunking is the default strategy that automatically identifies document structure and splits accordingly.

How It Works​

  • Identifies paragraphs, headers, lists, and other structures
  • Maintains semantic integrity
  • Automatically adjusts chunk size

Best For​

  • Structured documents (technical docs, reports)
  • Mixed content documents
  • Most general use cases

Configuration Parameters​

ParameterDescriptionDefault
chunk_sizeTarget chunk size (characters)500
chunk_overlapOverlap between chunks50

πŸ“ Sentence-based Chunking​

Sentence-based chunking splits documents by sentence boundaries, suitable for scenarios requiring precise retrieval.

How It Works​

  • Identifies sentence boundaries (periods, question marks, exclamation marks)
  • Combines adjacent sentences into chunks
  • Maintains sentence integrity

Best For​

  • FAQ documents
  • Q&A content
  • Scenarios requiring precise matching

Configuration Parameters​

ParameterDescriptionDefault
separatorSentence separators.!?
buffer_sizeSentence buffer count1

πŸ”— Semantic Chunking​

Semantic chunking splits based on content semantic similarity, suitable for complex documents.

How It Works​

  • Calculates semantic similarity between adjacent text
  • Splits at semantic change points
  • Maintains topic coherence

Best For​

  • Long articles
  • Documents with diverse topics
  • Scenarios requiring context coherence

Configuration Parameters​

ParameterDescriptionDefault
breakpoint_thresholdSemantic breakpoint threshold0.5
buffer_sizeContext buffer size1

πŸ“‘ Excel Files​

Excel files (.xlsx) are always chunked by whole rows, regardless of the selected strategy: a row is the smallest semantic unit of a table and is never cut in half.

How It Works​

  • Each worksheet is chunked independently, preserving the sheet name and physical row/column coordinates (newlines in sheet names are collapsed to spaces so they cannot forge row content)
  • Multiple complete rows are packed into one chunk up to the chunk size limit
  • A row exceeding the limit is split by cells; a single overlong cell value is split into coordinate-bearing fragments that can be fully reconstructed
  • Character overlap (chunk_overlap) does not apply to Excel
  • Formula cells index both the formula expression and its cached result (when the workbook was saved with one), so querying by either matches

Notes​

  • Semantic chunking does not apply to Excel; files fall back to row-wise chunking
  • chunk_size bounds both text forms of each chunk β€” the embedded text (readable rows plus the Worksheet context line) and the display text (canonical rows plus the sheet header); each prefix is paid out of the chunk size instead of being added on top (overlong sheet names are truncated with an ellipsis; real files are bounded by Excel's 31-character worksheet name limit). File-level origin is not part of the embedded text β€” it is carried by the file title in retrieval results and by document-scope filters
  • Row data outranks context: near the minimum chunk budget, if keeping the sheet prefix would starve a cell fragment of room, the prefix is dropped to preserve data completeness (an empty sheet name also emits no prefix)
  • With hierarchical chunking, both parent and child chunks pack whole rows
  • Chunking follows the file's real content: whatever file type the knowledge base is configured with, files that are actually Excel are always chunked by rows. Workbooks with corrupted style data are retried once with styles skipped; any other unparseable corrupted workbook fails indexing explicitly rather than degrading silently

βš™οΈ General Configuration​

Chunk Size​

Chunk size affects retrieval precision and recall:

SizeProsCons
Small (200-300)Precise matchingMay lose context
Medium (400-600)Balance precision and contextGeneral choice
Large (800-1000)Preserve more contextMay include irrelevant content

Chunk Overlap​

Overlap prevents important information from being split:

  • No overlap (0): Independent chunks, saves storage
  • Small overlap (20-50): Basic continuity
  • Large overlap (100+): Strong context preservation

πŸ’‘ Selection Recommendations​

By Document Type​

Document TypeRecommended StrategyReason
Technical docsSmart ChunkingPreserve structure
FAQSentence-basedPrecise Q&A matching
Long articlesSemanticMaintain topic coherence
Code docsSmart ChunkingIdentify code blocks

By Use Case​

ScenarioRecommended Configuration
Precise Q&ASentence-based + small chunks
Knowledge retrievalSmart Chunking + medium chunks
Context understandingSemantic + large chunks

πŸ”„ Re-chunking​

If retrieval results are unsatisfactory, you can re-chunk:

  1. Go to the knowledge base document list
  2. Select documents to re-chunk
  3. Click Re-index
  4. Choose new chunking strategy and parameters
  5. Confirm reprocessing

Re-chunking will delete old chunks and create new ones.

πŸ”— Related Documentation​