Module 1: How LLMs Actually Chunk Content

Why "sections" are not real chunks.

This module explains how LLMs actually break content into chunks, and why visual structure does not determine chunk boundaries. See also: Core Concepts

• Token-based chunking: LLMs chunk based on token limits, not visual structure
• DOM boundary heuristics: HTML structure influences but does not control chunking
• Sentence vs semantic chunking: Chunks may split mid-sentence or combine multiple sentences
• Overlap windows: Chunks overlap to preserve context, but overlap is not guaranteed
• Truncation bias: Long sentences or paragraphs are truncated, losing information

LLMs do not respect:

• paragraphs
• headings
• visual sections

They chunk based on:

• token limits
• punctuation density
• semantic similarity
• transformer attention patterns

If a fact requires previous text, it is already broken.

Design every chunk to be self-contained. Never assume surrounding context will be preserved.

Task: Select any existing webpage (yours or a competitor's). Copy three consecutive paragraphs. Split them into extraction-sized chunks as an LLM would, ignoring visual structure.

Constraint: Each chunk must be token-limited (approximately 512 tokens). Split at semantic boundaries, not sentence boundaries.

What success looks like: You produce a list of chunks where each chunk can be read in isolation without losing meaning. If a chunk requires previous context to be understood, you've identified a chunking failure.

This task is optional. No submission required. No validation. Use it to convert theory into applied thinking.