Chunking Strategies

When you add documents to a Knowledge Base, Tensoras splits them into chunks before generating embeddings and indexing. The chunking strategy determines how text is divided, which directly affects retrieval quality. Tensoras supports four chunking strategies.

Strategies at a Glance

Semantic Chunking (Default)

Semantic chunking splits text at natural topic boundaries by analyzing embedding similarity between consecutive sentences. When the similarity drops below a threshold, a new chunk starts. This produces chunks that are topically coherent, which improves retrieval relevance.

This is the default and recommended strategy for most use cases.

from tensoras import Tensoras
 
client = Tensoras(api_key="tns_your_key_here")
 
kb = client.knowledge_bases.create(
    name="Research Papers",
    embedding_model="bge-large-en-v1.5",
    chunking_strategy={
        "type": "semantic",
    },
)

Semantic chunking produces variable-length chunks. Typical chunks range from 200 to 1,500 tokens depending on the document structure.

Fixed Size Chunking

Fixed size chunking splits text into chunks of a specified character count with an optional overlap between consecutive chunks. Overlap helps preserve context at chunk boundaries.

kb = client.knowledge_bases.create(
    name="Legal Contracts",
    embedding_model="bge-large-en-v1.5",
    chunking_strategy={
        "type": "fixed_size",
        "chunk_size": 1000,       # characters per chunk
        "chunk_overlap": 200,     # overlap between consecutive chunks
    },
)

When to use: When you need predictable, uniform chunk sizes — for example, when working with highly uniform documents like legal contracts or log files.

Recursive Character Chunking

Recursive character chunking splits text using a hierarchy of separators, trying the most meaningful separator first and falling back to less meaningful ones:

1. \n\n — paragraph breaks
2. \n — line breaks
3. . — sentence boundaries
4. — word boundaries

This preserves document structure (paragraphs, sentences) while keeping chunks under the size limit.

kb = client.knowledge_bases.create(
    name="Technical Docs",
    embedding_model="bge-large-en-v1.5",
    chunking_strategy={
        "type": "recursive",
        "chunk_size": 1000,
        "chunk_overlap": 200,
    },
)

When to use: When documents have clear structural markers (paragraphs, sections) and you want chunks that respect those boundaries while staying under a size limit.

Hierarchical Chunking

Hierarchical chunking creates parent-child chunk relationships at multiple levels of granularity. Small child chunks are used for precise retrieval, while their larger parent chunks provide broader context when needed.

kb = client.knowledge_bases.create(
    name="Knowledge Base",
    embedding_model="bge-large-en-v1.5",
    chunking_strategy={
        "type": "hierarchical",
        "parent_chunk_size": 2000,
        "child_chunk_size": 500,
        "child_chunk_overlap": 100,
    },
)

During retrieval, Tensoras searches over child chunks for precision, then optionally expands to the parent chunk to provide the model with more surrounding context.

When to use: When your documents contain information at different levels of detail, or when you find that small chunks retrieve well but lack sufficient context for the model to answer accurately.

Configuration Reference

The chunking_strategy object is passed when creating a Knowledge Base:

{
  "chunking_strategy": {
    "type": "semantic | fixed_size | recursive | hierarchical",
    "chunk_size": 1000,
    "chunk_overlap": 200,
    "parent_chunk_size": 2000,
    "child_chunk_size": 500,
    "child_chunk_overlap": 100
  }
}

Choosing a Strategy

• Start with semantic chunking. It works well for most document types without tuning.
• Use fixed size when you need uniform chunks and your documents lack clear structural markers.
• Use recursive when documents have well-defined paragraph and section structure.
• Use hierarchical when retrieval precision matters and you need the model to have broader context around the matched passage.

Related

• RAG Overview — end-to-end RAG pipeline
• Hybrid Search — how chunks are searched and ranked
• Knowledge Bases API — create and configure Knowledge Bases