Late Chunking

Embed first, then split for better chunk embeddings

Late chunking embeds the entire document first, then splits. This produces chunks with better individual embeddings.

Quick Start

Agent with Late Chunking

from praisonaiagents import Agent

agent = Agent(
    instructions="Answer questions with high precision.",
    knowledge={
        "sources": ["technical_docs/"],
        "chunker": {
            "type": "late",
            "chunk_size": 512,
            "embedding_model": "all-MiniLM-L6-v2"
        }
    }
)

response = agent.start("Explain the architecture")

from praisonaiagents.knowledge.chunking import Chunking

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

chunks = chunker.chunk("Your document content...")

When to Use

• High-precision retrieval needed
• Quality matters more than speed
• Complex technical documents
• Semantic similarity search critical

Parameters

Parameter	Type	Default	Description
chunk_size	int	512	Max tokens per chunk
embedding_model	str	auto	Embedding model

How It Works

Traditional chunking: Split → Embed each chunk Late chunking: Embed full doc → Split with context awareness

This preserves document-level context in each chunk’s embedding.