Semantic Chunking - PraisonAI

Uses embeddings to split text at semantic boundaries. Groups related content together for better retrieval.

Quick Start

from praisonaiagents import Agent

agent = Agent(
    instructions="Answer questions from research papers.",
    knowledge={
        "sources": ["papers/"],
        "chunker": {
            "type": "semantic",
            "chunk_size": 512,
            "embedding_model": "all-MiniLM-L6-v2"
        }
    }
)

response = agent.start("What methodology did they use?")

from praisonaiagents.knowledge.chunking import Chunking

chunker = Chunking(
    chunker_type="semantic",
    chunk_size=512,
    embedding_model="all-MiniLM-L6-v2"
)

chunks = chunker.chunk("Your research paper content here...")
for chunk in chunks:
    print(chunk.text)

When to Use

Good For

Research papers
Topic-dense content
Multi-subject documents
Quality over speed

Consider Alternatives

Speed-critical pipelines
Uniform chunk sizes needed
Simple structured content
Very short documents

Parameters

Parameter	Type	Default	Description
`chunk_size`	int	512	Max tokens per chunk
`embedding_model`	str	auto	Model for semantic similarity

Examples

Research Analysis

agent = Agent(
    instructions="Analyze academic papers.",
    knowledge={
        "sources": ["research/"],
        "chunker": {
            "type": "semantic",
            "chunk_size": 512,
            "embedding_model": "all-MiniLM-L6-v2"
        }
    }
)

Knowledge Base

agent = Agent(
    instructions="Answer from knowledge base.",
    knowledge={
        "sources": ["wiki/", "faq.txt"],
        "chunker": {
            "type": "semantic",
            "chunk_size": 256  # Smaller for precise retrieval
        }
    }
)

How It Works

Semantic chunking:

Splits document into sentences
Generates embeddings for each sentence
Groups consecutive similar sentences
Creates new chunk when topic changes

Performance Note

Semantic chunking requires computing embeddings and is slower than token/sentence chunking. Use for quality-sensitive applications where retrieval accuracy matters more than speed.

Embedding Models

The default embedding model is all-MiniLM-L6-v2. You can use any model supported by the chonkie library:

knowledge={
    "sources": ["docs/"],
    "chunker": {
        "type": "semantic",
        "embedding_model": "all-MiniLM-L6-v2"  # Default
        # Or: "text-embedding-3-small", etc.
    }
}

Recursive Chunking

SDPM Chunking

​Quick Start

​When to Use

Good For

Consider Alternatives

​Parameters

​Examples

​Research Analysis

​Knowledge Base

​How It Works

​Performance Note

​Embedding Models

Quick Start

When to Use

Parameters

Examples

Research Analysis

Knowledge Base

How It Works

Performance Note

Embedding Models