Chunking Strategy: Sliding Window

Overview

RAGv2 uses a “Sliding Window” strategy. This prevents text from being split awkwardly mid-sentence. It moves back a defined “overlap” amount before starting the next chunk.

Flow / Behavior

graph TD
    Text["📜 Raw Document Text"] --> Parent["✂️ Split into Parent Chunks"]
    Parent --> Child["✂️ Split Parents into Child Chunks"]
    Child --> Vector["🔢 Convert Children to Vectors"]

Configuration

Parameter	Type	Default	Description
PARENT_CHUNK_SIZE	int	2000	Context given to the AI. Contains full paragraphs.
PARENT_OVERLAP	int	250	Overlap for Parents to ensure smooth transitions.
CHILD_CHUNK_SIZE	int	250	Specific target size for FAISS vectors.
CHILD_OVERLAP	int	75	Overlap for Children.
TOP_K	int	4	Number of Parent chunks retrieved.

Related Components

• system-architecture
• data-pipeline

---

Last Updated: 2026-05-01